Explainable artificial intelligence

Cooperation between agents – in this case, algorithms and humans – depends on trust. If humans are to accept algorithmic prescriptions, they need to trust them. Incompleteness in formal trust criteria is a barrier to optimization. Transparency, interpretability, and explainability are intermediate goals on the road to these more comprehensive trust criteria.{{cite conference |url=http://docs.mipro-proceedings.com/dsdc/dsdc_11_4754.pdf |title=Explainable Artificial Intelligence: A Survey |last1=Dosilovic |first1=Filip |last2=Brcic |first2=Mario |last3=Hlupic |first3=Nikica |date=2018-05-25 |book-title=MIPRO 2018 - 41st International Convention Proceedings |pages=210–215 |location=Opatija, Croatia |conference=MIPRO 2018 |doi=10.23919/MIPRO.2018.8400040 |isbn=978-953-233-095-3 |access-date=2018-12-09 |archive-date=2018-12-10 |archive-url=https://web.archive.org/web/20181210110820/http://docs.mipro-proceedings.com/dsdc/dsdc_11_4754.pdf |url-status=dead }} This is particularly relevant in medicine,{{Cite journal |last1=Bernal |first1=Jose |last2=Mazo |first2=Claudia |date=2022-10-11 |title=Transparency of Artificial Intelligence in Healthcare: Insights from Professionals in Computing and Healthcare Worldwide |journal=Applied Sciences |language=en |volume=12 |issue=20 |pages=10228 |doi=10.3390/app122010228 |issn=2076-3417|doi-access=free }} especially with clinical decision support systems (CDSS), in which medical professionals should be able to understand how and why a machine-based decision was made in order to trust the decision and augment their decision-making process.{{Cite journal |last1=Antoniadi |first1=Anna Markella |last2=Du |first2=Yuhan |last3=Guendouz |first3=Yasmine |last4=Wei |first4=Lan |last5=Mazo |first5=Claudia |last6=Becker |first6=Brett A. |last7=Mooney |first7=Catherine |date=January 2021 |title=Current Challenges and Future Opportunities for XAI in Machine Learning-Based Clinical Decision Support Systems: A Systematic Review |journal=Applied Sciences |language=en |volume=11 |issue=11 |pages=5088 |doi=10.3390/app11115088 |issn=2076-3417|doi-access=free }}

AI systems sometimes learn undesirable tricks that do an optimal job of satisfying explicit pre-programmed goals on the training data but do not reflect the more nuanced implicit desires of the human system designers or the full complexity of the domain data. For example, a 2017 system tasked with image recognition learned to "cheat" by looking for a copyright tag that happened to be associated with horse pictures rather than learning how to tell if a horse was actually pictured. In another 2017 system, a supervised learning AI tasked with grasping items in a virtual world learned to cheat by placing its manipulator between the object and the viewer in a way such that it falsely appeared to be grasping the object.{{cite news|title=DeepMind Has Simple Tests That Might Prevent Elon Musk's AI Apocalypse|url=https://www.bloomberg.com/news/articles/2017-12-11/deepmind-has-simple-tests-that-might-prevent-elon-musk-s-ai-apocalypse|access-date=30 January 2018|work=Bloomberg.com|date=11 December 2017|language=en}}{{cite news|title=Learning from Human Preferences|url=https://blog.openai.com/deep-reinforcement-learning-from-human-preferences/|access-date=30 January 2018|work=OpenAI Blog|date=13 June 2017}}

One transparency project, the DARPA XAI program, aims to produce "glass box" models that are explainable to a "human-in-the-loop" without greatly sacrificing AI performance. Human users of such a system can understand the AI's cognition (both in real-time and after the fact) and can determine whether to trust the AI.{{cite web|title=Explainable Artificial Intelligence (XAI)|url=https://www.darpa.mil/program/explainable-artificial-intelligence|website=DARPA|access-date=17 July 2017}} Other applications of XAI are knowledge extraction from black-box models and model comparisons.{{cite journal|last=Biecek|first=Przemyslaw|title= DALEX: explainers for complex predictive models|journal=Journal of Machine Learning Research|volume=19|pages=1–5|arxiv=1806.08915|date=23 June 2018}} In the context of monitoring systems for ethical and socio-legal compliance, the term "glass box" is commonly used to refer to tools that track the inputs and outputs of the system in question, and provide value-based explanations for their behavior. These tools aim to ensure that the system operates in accordance with ethical and legal standards, and that its decision-making processes are transparent and accountable. The term "glass box" is often used in contrast to "black box" systems, which lack transparency and can be more difficult to monitor and regulate.Rai, Arun. "Explainable AI: From black box to glass box." Journal of the Academy of Marketing Science 48 (2020): 137-141.

The term is also used to name a voice assistant that produces counterfactual statements as explanations.{{cite book|last1=Sokol|first1=Kacper|last2=Flach|first2=Peter|title=Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence|chapter=Glass-Box: Explaining AI Decisions With Counterfactual Statements Through Conversation With a Voice-enabled Virtual Assistant|year=2018|pages=5868–5870|doi=10.24963/ijcai.2018/865|isbn=9780999241127|s2cid=51608978}}

There is a subtle difference between the terms explainability and interpretability in the context of AI.{{cite journal |last1=Broniatowski |first1=David A. |title=Psychological Foundations of Explainability and Interpretability in Artificial Intelligence |journal=NIST Pubs |date=2021 |url=https://www.nist.gov/publications/psychological-foundations-explainability-and-interpretability-artificial-intelligence}}

Some explainability techniques don't involve understanding how the model works, and may work across various AI systems. Treating the model as a black box and analyzing how marginal changes to the inputs affect the result sometimes provides a sufficient explanation.

Explainability is useful for ensuring that AI models are not making decisions based on irrelevant or otherwise unfair criteria. For classification and regression models, several popular techniques exist:

• Partial dependency plots show the marginal effect of an input feature on the predicted outcome.
• SHAP (SHapley Additive exPlanations) enables visualization of the contribution of each input feature to the output. It works by calculating Shapley values, which measure the average marginal contribution of a feature across all possible combinations of features.{{Cite web |last=Verma |first=Yugesh |date=2021-12-25 |title=Complete Guide to SHAP - SHAPley Additive exPlanations for Practitioners |url=https://analyticsindiamag.com/a-complete-guide-to-shap-shapley-additive-explanations-for-practitioners/ |access-date=2024-07-10 |website=Analytics India Magazine |language=en-US}}
• Feature importance estimates how important a feature is for the model. It is usually done using permutation importance, which measures the performance decrease when it the feature value randomly shuffled across all samples.
• LIME approximates locally a model's outputs with a simpler, interpretable model.{{Cite web |last=Rothman |first=Denis |date=2020-10-07 |title=Exploring LIME Explanations and the Mathematics Behind It |url=https://www.codemotion.com/magazine/ai-ml/lime-explainable-ai/ |access-date=2024-07-10 |website=Codemotion Magazine |language=en-US}}
• Multitask learning provides a large number of outputs in addition to the target classification. These other outputs can help developers deduce what the network has learned.{{Cite book |last=Christian |first=Brian |title=The Alignment Problem: Machine learning and human values |date=2020 |publisher=W. W. Norton & Company |isbn=978-0-393-86833-3 |chapter=TELL ME EVERYTHING: MULTITASK NETS}}

For images, saliency maps highlight the parts of an image that most influenced the result.{{Cite web |last=Sharma |first=Abhishek |date=2018-07-11 |title=What Are Saliency Maps In Deep Learning? |url=https://analyticsindiamag.com/what-are-saliency-maps-in-deep-learning/ |access-date=2024-07-10 |website=Analytics India Magazine |language=en-US}}

Systems that are expert or knowledge based are software systems that are made by experts. This system consists of a knowledge based encoding for the domain knowledge. This system is usually modeled as production rules, and someone uses this knowledge base which the user can question the system for knowledge. In expert systems, the language and explanations are understood with an explanation for the reasoning or a problem solving activity.

However, these techniques are not very suitable for language models like generative pretrained transformers. Since these models generate language, they can provide an explanation, but which may not be reliable. Other techniques include attention analysis (examining how the model focuses on different parts of the input), probing methods (testing what information is captured in the model's representations), causal tracing (tracing the flow of information through the model) and circuit discovery (identifying specific subnetworks responsible for certain behaviors). Explainability research in this area overlaps significantly with interpretability and alignment research.{{cite arXiv |last1=Luo |first1=Haoyan |title=From Understanding to Utilization: A Survey on Explainability for Large Language Models |date=2024-02-21 |eprint=2401.12874 |last2=Specia |first2=Lucia|class=cs.CL }}

{{Main|Mechanistic interpretability}}

File:Grokking modular addition.jpg is an example of phenomenon studied in interpretability. It involves a model that initially memorizes all the answers (overfitting), but later adopts an algorithm that generalizes to unseen data.{{Cite web |last=Ananthaswamy |first=Anil |date=2024-04-12 |title=How Do Machines ‘Grok’ Data? |url=https://www.quantamagazine.org/how-do-machines-grok-data-20240412/ |access-date=2025-01-21 |website=Quanta Magazine |language=en}}]]

Scholars sometimes use the term "mechanistic interpretability" to refer to the process of reverse-engineering artificial neural networks to understand their internal decision-making mechanisms and components, similar to how one might analyze a complex machine or computer program.{{Cite web |last=Olah |first=Chris |date=June 27, 2022 |title=Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases |url=https://www.transformer-circuits.pub/2022/mech-interp-essay |access-date=2024-07-10 |website=www.transformer-circuits.pub}}

Interpretability research often focuses on generative pretrained transformers. It is particularly relevant for AI safety and alignment, as it may enable to identify signs of undesired behaviors such as sycophancy, deceptiveness or bias, and to better steer AI models.{{Cite web |last=Mittal |first=Aayush |date=2024-06-17 |title=Understanding Sparse Autoencoders, GPT-4 & Claude 3 : An In-Depth Technical Exploration |url=https://www.unite.ai/understanding-sparse-autoencoders-gpt-4-claude-3-an-in-depth-technical-exploration/ |access-date=2024-07-10 |website=Unite.AI |language=en-US}}

Studying the interpretability of the most advanced foundation models often involves searching for an automated way to identify "features" in generative pretrained transformers. In a neural network, a feature is a pattern of neuron activations that corresponds to a concept. A compute-intensive technique called "dictionary learning" makes it possible to identify features to some degree. Enhancing the ability to identify and edit features is expected to significantly improve the safety of frontier AI models.{{Cite web |last=Ropek |first=Lucas |date=2024-05-21 |title=New Anthropic Research Sheds Light on AI's 'Black Box' |url=https://gizmodo.com/new-anthropic-research-sheds-light-on-ais-black-box-1851491333 |access-date=2024-05-23 |website=Gizmodo |language=en}}{{Cite magazine |last=Perrigo |first=Billy |date=2024-05-21 |title=Artificial Intelligence Is a 'Black Box.' Maybe Not For Long |url=https://time.com/6980210/anthropic-interpretability-ai-safety-research/ |access-date=2024-05-24 |magazine=Time |language=en}}

For convolutional neural networks, DeepDream can generate images that strongly activate a particular neuron, providing a visual hint about what the neuron is trained to identify.{{Cite magazine |last=Barber |first=Gregory |title=Inside the 'Black Box' of a Neural Network |url=https://www.wired.com/story/inside-black-box-of-neural-network/ |access-date=2024-07-10 |magazine=Wired |language=en-US |issn=1059-1028}}

During the 1970s to 1990s, symbolic reasoning systems, such as MYCIN,{{cite journal |last1=Fagan |first1=L. M. |last2=Shortliffe |first2=E. H. |last3=Buchanan |first3=B. G. |year=1980 |title=Computer-based medical decision making: from MYCIN to VM |journal=Automedica |volume=3 |issue=2 |pages=97–108}} GUIDON,{{Cite book| publisher = The MIT Press| last = Clancey| first = William| title = Knowledge-Based Tutoring: The GUIDON Program| location = Cambridge, Massachusetts| date = 1987}} SOPHIE,{{Cite book| publisher = Academic Press| isbn = 0-12-648680-8| last1 = Brown| first1 = John S.| last2 = Burton| first2 = R. R.| last3 = De Kleer| first3 = Johan| title = Intelligent Tutoring Systems| chapter = Pedagogical, natural language, and knowledge engineering techniques in SOPHIE I, II, and III| date = 1982}} and PROTOS{{Cite book| publisher = Morgan Kaufmann Publishers Inc.| isbn = 1-55860-119-8| volume = 3| pages = 112–139| last1 = Bareiss| first1 = Ray| last2 = Porter| first2 = Bruce| last3 = Weir| first3 = Craig| last4 = Holte| first4 = Robert| title = Machine Learning| chapter = Protos: An Exemplar-Based Learning Apprentice| date = 1990}}{{Cite book |last=Bareiss |first=Ray |title=Exemplar-Based Knowledge Acquisition: A Unified Approach to Concept Representation, Classification, and Learning |publisher=Academic Press |year=1989 |series=Perspectives in Artificial Intelligence}} could represent, reason about, and explain their reasoning for diagnostic, instructional, or machine-learning (explanation-based learning) purposes. MYCIN, developed in the early 1970s as a research prototype for diagnosing bacteremia infections of the bloodstream, could explain{{cite book |last1=Van Lent |first1=M. |last2=Fisher |first2=W. |last3=Mancuso |first3=M. |date=July 2004 |chapter=An explainable artificial intelligence system for small-unit tactical behavior |title=Proceedings of the National Conference on Artificial Intelligence |pages=900–907 |location=San Jose, CA |publisher=AAAI Press |isbn=0262511835}} which of its hand-coded rules contributed to a diagnosis in a specific case. Research in intelligent tutoring systems resulted in developing systems such as SOPHIE that could act as an "articulate expert", explaining problem-solving strategy at a level the student could understand, so they would know what action to take next. For instance, SOPHIE could explain the qualitative reasoning behind its electronics troubleshooting, even though it ultimately relied on the SPICE circuit simulator. Similarly, GUIDON added tutorial rules to supplement MYCIN's domain-level rules so it could explain the strategy for medical diagnosis. Symbolic approaches to machine learning relying on explanation-based learning, such as PROTOS, made use of explicit representations of explanations expressed in a dedicated explanation language, both to explain their actions and to acquire new knowledge.

In the 1980s through the early 1990s, truth maintenance systems (TMS) extended the capabilities of causal-reasoning, rule-based, and logic-based inference systems.{{Cite book| edition = Second| publisher = Prentice Hall, Pearson Education| isbn = 0-13-790395-2| last1 = Russell| first1 = Stuart| last2 = Norvig| first2 = Peter| title = Artificial Intelligence: A Modern Approach| location = Upper Saddle River, New Jersey| series = Prentice Hall Series in Artificial Intelligence| date = 2003}}{{rp|360–362}} A TMS explicitly tracks alternate lines of reasoning, justifications for conclusions, and lines of reasoning that lead to contradictions, allowing future reasoning to avoid these dead ends. To provide an explanation, they trace reasoning from conclusions to assumptions through rule operations or logical inferences, allowing explanations to be generated from the reasoning traces. As an example, consider a rule-based problem solver with just a few rules about Socrates that concludes he has died from poison:

{{Blockquote

|text=By just tracing through the dependency structure the problem solver can construct the following explanation: "Socrates died because he was mortal and drank poison, and all mortals die when they drink poison. Socrates was mortal because he was a man and all men are mortal. Socrates drank poison because he held dissident beliefs, the government was conservative, and those holding conservative dissident beliefs under conservative governments must drink poison."{{Cite book| publisher = The MIT Press| isbn = 0-262-06157-0| last1 = Forbus| first1 = Kenneth| last2 = De Kleer| first2 = Johan| title = Building Problem Solvers| location = Cambridge, Massachusetts| date = 1993}}{{rp|164–165}}

}}

By the 1990s researchers began studying whether it is possible to meaningfully extract the non-hand-coded rules being generated by opaque trained neural networks.{{Cite journal|last1=Tickle|first1=A. B.|last2=Andrews|first2=R.|last3=Golea|first3=M.|last4=Diederich|first4=J.|date=November 1998|title=The truth will come to light: directions and challenges in extracting the knowledge embedded within trained artificial neural networks|journal=IEEE Transactions on Neural Networks|volume=9|issue=6|pages=1057–1068|doi=10.1109/72.728352|pmid=18255792|s2cid=11111578 |issn=1045-9227}} Researchers in clinical expert systems creating{{clarify|reason=who is creating? the researchers or the expert systems?|date=April 2023}} neural network-powered decision support for clinicians sought to develop dynamic explanations that allow these technologies to be more trusted and trustworthy in practice. In the 2010s public concerns about racial and other bias in the use of AI for criminal sentencing decisions and findings of creditworthiness may have led to increased demand for transparent artificial intelligence. As a result, many academics and organizations are developing tools to help detect bias in their systems.{{cite news|url=https://www.bloomberg.com/news/articles/2018-06-13/accenture-unveils-tool-to-help-companies-insure-their-ai-is-fair|title=Accenture Unveils Tool to Help Companies Insure Their AI Is Fair|date=June 2018|work=Bloomberg.com|access-date=5 August 2018|language=en}}

Marvin Minsky et al. raised the issue that AI can function as a form of surveillance, with the biases inherent in surveillance, suggesting HI (Humanistic Intelligence) as a way to create a more fair and balanced "human-in-the-loop" AI.Minsky, et al., "The Society of Intelligent Veillance" IEEE ISTAS2013, pages 13-17.

Explainable AI has been recently a new topic researched amongst the context of modern deep learning. Modern complex AI techniques, such as deep learning, are naturally opaque.{{cite magazine|last1=Mukherjee|first1=Siddhartha|title=A.I. Versus M.D.|url=https://www.newyorker.com/magazine/2017/04/03/ai-versus-md|access-date=30 January 2018|magazine=The New Yorker|date=27 March 2017}} To address this issue, methods have been developed to make new models more explainable and interpretable.{{Cite journal|date=2020-07-08|title=Interpretable neural networks based on continuous-valued logic and multicriteria decision operators|journal=Knowledge-Based Systems|language=en|volume=199|pages=105972|doi=10.1016/j.knosys.2020.105972 |arxiv=1910.02486 |issn=0950-7051|doi-access=free|last1=Csiszár|first1=Orsolya|last2=Csiszár|first2=Gábor|last3=Dombi|first3=József}}{{cite arXiv|last1=Doshi-Velez|first1=Finale|last2=Kim|first2=Been|date=2017-02-27|title=Towards A Rigorous Science of Interpretable Machine Learning|eprint=1702.08608|class=stat.ML}}{{Cite arXiv |last=Abdollahi, Behnoush, and Olfa Nasraoui.|title=Explainable Restricted Boltzmann Machines for Collaborative Filtering.|eprint=1606.07129|class=stat.ML|year=2016}}{{Cite book|last1=Dombi|first1=József|last2=Csiszár|first2=Orsolya|series=Studies in Fuzziness and Soft Computing |date=2021|title=Explainable Neural Networks Based on Fuzzy Logic and Multi-criteria Decision Tools|url=https://link.springer.com/book/10.1007/978-3-030-72280-7|volume=408|language=en-gb|doi=10.1007/978-3-030-72280-7|isbn=978-3-030-72279-1|s2cid=233486978|issn=1434-9922}} This includes layerwise relevance propagation (LRP), a technique for determining which features in a particular input vector contribute most strongly to a neural network's output.{{cite journal|last1=Bach|first1=Sebastian|last2=Binder|first2=Alexander|last3=Montavon|first3=Grégoire|last4=Klauschen|first4=Frederick|last5=Müller|first5=Klaus-Robert|author-link5=Klaus-Robert Müller|last6=Samek|first6=Wojciech|date=2015-07-10|editor-last=Suarez|editor-first=Oscar Deniz|title=On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation|journal=PLOS ONE|volume=10|issue=7|page=e0130140|bibcode=2015PLoSO..1030140B|doi=10.1371/journal.pone.0130140|issn=1932-6203|pmc=4498753|pmid=26161953|doi-access=free}}{{cite news|url=https://www.theguardian.com/science/2017/nov/05/computer-says-no-why-making-ais-fair-accountable-and-transparent-is-crucial|title=Computer says no: why making AIs fair, accountable and transparent is crucial|last1=Sample|first1=Ian|date=5 November 2017|work=The Guardian|access-date=5 August 2018|language=en}} Other techniques explain some particular prediction made by a (nonlinear) black-box model, a goal referred to as "local interpretability".{{Cite journal|last1=Martens|first1=David|last2=Provost|first2=Foster|title=Explaining data-driven document classifications|url=http://pages.stern.nyu.edu/~fprovost/Papers/MartensProvost_Explaining.pdf|journal=MIS Quarterly|year=2014|volume=38|pages=73–99|doi=10.25300/MISQ/2014/38.1.04|s2cid=14238842}}{{Cite journal|title="Why Should I Trust You?" {{!}} Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining|language=EN|doi=10.1145/2939672.2939778|s2cid=13029170}}{{Citation|last1=Lundberg|first1=Scott M|title=A Unified Approach to Interpreting Model Predictions|date=2017|url=http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf|work=Advances in Neural Information Processing Systems 30|pages=4765–4774|editor-last=Guyon|editor-first=I.|publisher=Curran Associates, Inc.|access-date=2020-03-13|last2=Lee|first2=Su-In|editor2-last=Luxburg|editor2-first=U. V.|editor3-last=Bengio|editor3-first=S.|editor4-last=Wallach|editor4-first=H.|arxiv=1705.07874}}{{Cite journal|last1=Carter|first1=Brandon|last2=Mueller|first2=Jonas|last3=Jain|first3=Siddhartha|last4=Gifford|first4=David|date=2019-04-11|title=What made you do this? Understanding black-box decisions with sufficient input subsets|url=http://proceedings.mlr.press/v89/carter19a.html|journal=The 22nd International Conference on Artificial Intelligence and Statistics|language=en|pages=567–576}}{{Cite journal|last1=Shrikumar|first1=Avanti|last2=Greenside|first2=Peyton|last3=Kundaje|first3=Anshul|date=2017-07-17|title=Learning Important Features Through Propagating Activation Differences|url=http://proceedings.mlr.press/v70/shrikumar17a.html|journal=International Conference on Machine Learning|language=en|pages=3145–3153}}{{Cite journal|url=https://dl.acm.org/doi/abs/10.5555/3305890.3306024|title=Axiomatic attribution for deep networks {{!}} Proceedings of the 34th International Conference on Machine Learning - Volume 70|website=dl.acm.org|series=Icml'17|date=6 August 2017|pages=3319–3328|language=EN|access-date=2020-03-13}} We still today cannot explain the output of today's DNNs without the new explanatory mechanisms, we also can't by the neural network, or external explanatory components {{Citation |last1=Xu |first1=Feiyu |title=Explainable AI: A Brief Survey on History, Research Areas, Approaches and Challenges |date=2019 |work=Natural Language Processing and Chinese Computing |volume=11839 |pages=563–574 |editor-last=Tang |editor-first=Jie |url=http://link.springer.com/10.1007/978-3-030-32236-6_51 |access-date=2024-12-03 |place=Cham |publisher=Springer International Publishing |language=en |doi=10.1007/978-3-030-32236-6_51 |isbn=978-3-030-32235-9 |last2=Uszkoreit |first2=Hans |last3=Du |first3=Yangzhou |last4=Fan |first4=Wei |last5=Zhao |first5=Dongyan |last6=Zhu |first6=Jun |editor2-last=Kan |editor2-first=Min-Yen |editor3-last=Zhao |editor3-first=Dongyan |editor4-last=Li |editor4-first=Sujian}} There is also research on whether the concepts of local interpretability can be applied to a remote context, where a model is operated by a third-party.{{Cite journal|last1=Le Merrer|first1=Erwan|last2=Trédan|first2=Gilles|date=September 2020|title=Remote explainability faces the bouncer problem|url=https://www.nature.com/articles/s42256-020-0216-z|journal=Nature Machine Intelligence|language=en|volume=2|issue=9|pages=529–539|doi=10.1038/s42256-020-0216-z|issn=2522-5839|arxiv=1910.01432|s2cid=225207140}}{{Cite journal |last1=Aivodji |first1=Ulrich |last2=Arai |first2=Hiromi |last3=Fortineau |first3=Olivier |last4=Gambs |first4=Sébastien |last5=Hara |first5=Satoshi |last6=Tapp |first6=Alain |date=2019-05-24 |title=Fairwashing: the risk of rationalization |url=http://proceedings.mlr.press/v97/aivodji19a.html |journal=International Conference on Machine Learning |language=en |publisher=PMLR |pages=161–170 |arxiv=1901.09749}}

There has been work on making glass-box models which are more transparent to inspection.{{cite journal |last1=Singh |first1=Chandan |last2=Nasseri |first2=Keyan |last3=Tan |first3=Yan Shuo |last4=Tang |first4=Tiffany |last5=Yu |first5=Bin |title=imodels: a python package for fitting interpretable models |journal=Journal of Open Source Software |date=4 May 2021 |volume=6 |issue=61 |pages=3192 |doi=10.21105/joss.03192 |bibcode=2021JOSS....6.3192S |s2cid=235529515 |url=https://joss.theoj.org/papers/10.21105/joss.03192 |language=en |issn=2475-9066}} This includes decision trees,{{Cite journal|last1=Vidal|first1=Thibaut|last2=Schiffer|first2=Maximilian|date=2020|title=Born-Again Tree Ensembles|url=http://proceedings.mlr.press/v119/vidal20a.html|journal=International Conference on Machine Learning|language=en|publisher=PMLR|volume=119|pages=9743–9753|arxiv=2003.11132}} Bayesian networks, sparse linear models,{{cite journal |last1=Ustun |first1=Berk |last2=Rudin |first2=Cynthia |title=Supersparse linear integer models for optimized medical scoring systems |journal=Machine Learning |date=1 March 2016 |volume=102 |issue=3 |pages=349–391 |doi=10.1007/s10994-015-5528-6 |s2cid=207211836 |url=https://link.springer.com/article/10.1007/s10994-015-5528-6 |language=en |issn=1573-0565}} and more.Bostrom, N., & Yudkowsky, E. (2014). [https://intelligence.org/files/EthicsofAI.pdf The ethics of artificial intelligence]. The Cambridge Handbook of Artificial Intelligence, 316-334. The Association for Computing Machinery Conference on Fairness, Accountability, and Transparency (ACM FAccT) was established in 2018 to study transparency and explainability in the context of socio-technical systems, many of which include artificial intelligence.{{cite web | url=https://fatconference.org/ | title=FAT* Conference }}{{cite news |title=Computer programs recognise white men better than black women |url=https://www.economist.com/science-and-technology/2018/02/15/computer-programs-recognise-white-men-better-than-black-women |access-date=5 August 2018 |newspaper=The Economist |date=2018 |language=en}}

Some techniques allow visualisations of the inputs to which individual software neurons respond to most strongly. Several groups found that neurons can be aggregated into circuits that perform human-comprehensible functions, some of which reliably arise across different networks trained independently.{{cite journal |last1=Olah |first1=Chris |last2=Cammarata |first2=Nick |last3=Schubert |first3=Ludwig |last4=Goh |first4=Gabriel |last5=Petrov |first5=Michael |last6=Carter |first6=Shan |title=Zoom In: An Introduction to Circuits |journal=Distill |date=10 March 2020 |volume=5 |issue=3 |pages=e00024.001 |doi=10.23915/distill.00024.001 |url=https://distill.pub/2020/circuits/zoom-in/ |language=en |issn=2476-0757|doi-access=free }}{{cite journal |last1=Li |first1=Yixuan |last2=Yosinski |first2=Jason |last3=Clune |first3=Jeff |last4=Lipson |first4=Hod |last5=Hopcroft |first5=John |title=Convergent Learning: Do different neural networks learn the same representations? |journal=Feature Extraction: Modern Questions and Challenges |date=8 December 2015 |pages=196–212 |url=http://proceedings.mlr.press/v44/li15convergent.html |publisher=PMLR |language=en}}

There are various techniques to extract compressed representations of the features of given inputs, which can then be analysed by standard clustering techniques. Alternatively, networks can be trained to output linguistic explanations of their behaviour, which are then directly human-interpretable.{{cite book |last1=Hendricks |first1=Lisa Anne |last2=Akata |first2=Zeynep |last3=Rohrbach |first3=Marcus |last4=Donahue |first4=Jeff |last5=Schiele |first5=Bernt |last6=Darrell |first6=Trevor |title=Computer Vision – ECCV 2016 |chapter=Generating Visual Explanations |series=Lecture Notes in Computer Science |date=2016 |volume=9908 |pages=3–19 |doi=10.1007/978-3-319-46493-0_1 |chapter-url=https://link.springer.com/chapter/10.1007/978-3-319-46493-0_1 |publisher=Springer International Publishing |language=en|arxiv=1603.08507 |isbn=978-3-319-46492-3 |s2cid=12030503 }} Model behaviour can also be explained with reference to training data—for example, by evaluating which training inputs influenced a given behaviour the most.{{cite journal |last1=Koh |first1=Pang Wei |last2=Liang |first2=Percy |title=Understanding Black-box Predictions via Influence Functions |journal=International Conference on Machine Learning |date=17 July 2017 |pages=1885–1894 |url=http://proceedings.mlr.press/v70/koh17a.html |publisher=PMLR |arxiv=1703.04730 |language=en}}

The use of explainable artificial intelligence (XAI) in pain research, specifically in understanding the role of electrodermal activity for automated pain recognition: hand-crafted features and deep learning models in pain recognition, highlighting the insights that simple hand-crafted features can yield comparative performances to deep learning models and that both traditional feature engineering and deep feature learning approaches rely on simple characteristics of the input time-series data.{{Cite journal |last1=Gouverneur |first1=Philip |last2=Li |first2=Frédéric |last3=Shirahama |first3=Kimiaki |last4=Luebke |first4=Luisa |last5=Adamczyk |first5=Wacław M. |last6=Szikszay |first6=Tibor M. |last7=Luedtke |first7=Kerstin |last8=Grzegorzek |first8=Marcin |date=2023-02-09 |title=Explainable Artificial Intelligence (XAI) in Pain Research: Understanding the Role of Electrodermal Activity for Automated Pain Recognition |journal=Sensors |language=en |volume=23 |issue=4 |pages=1959 |doi=10.3390/s23041959 |issn=1424-8220 |pmc=9960387 |pmid=36850556 |bibcode=2023Senso..23.1959G |doi-access=free }}

As regulators, official bodies, and general users come to depend on AI-based dynamic systems, clearer accountability will be required for automated decision-making processes to ensure trust and transparency. The first global conference exclusively dedicated to this emerging discipline was the 2017 International Joint Conference on Artificial Intelligence: Workshop on Explainable Artificial Intelligence (XAI).{{cite web|title=IJCAI 2017 Workshop on Explainable Artificial Intelligence (XAI)|url=http://www.intelligentrobots.org/files/IJCAI2017/IJCAI-17_XAI_WS_Proceedings.pdf|website=Earthlink|publisher=IJCAI|access-date=17 July 2017|archive-date=4 April 2019|archive-url=https://web.archive.org/web/20190404131609/http://www.intelligentrobots.org/files/IJCAI2017/IJCAI-17_XAI_WS_Proceedings.pdf|url-status=dead}} It has evolved over the years, with various workshops organised and co-located to many other international conferences, and it has now a dedicated global event, "The world conference on eXplainable Artificial Intelligence", with its own proceedings.{{cite book |author= |date= 2023| title= Explainable Artificial Intelligence, First World Conference, xAI 2023, Lisbon, Portugal, July 26–28, 2023, Proceedings, Parts I/II/III |series= Communications in Computer and Information Science|volume= 1903|url= https://link.springer.com/book/10.1007/978-3-031-44070-0 |publisher=springer |doi= 10.1007/978-3-031-44070-0|isbn=978-3-031-44070-0}}{{cite book |author=|date= 2024| title= Explainable Artificial Intelligence, Second World Conference, xAI 2024, Valletta, Malta, July 17–19, 2024, Proceedings, Part I/II/III/IV |series= Communications in Computer and Information Science|volume= 2153|url=https://link.springer.com/book/10.1007/978-3-031-63787-2 |publisher=springer |doi= 10.1007/978-3-031-63787-2|isbn=978-3-031-63787-2}}

The European Union introduced a right to explanation in the General Data Protection Regulation (GDPR) to address potential problems stemming from the rising importance of algorithms. The implementation of the regulation began in 2018. However, the right to explanation in GDPR covers only the local aspect of interpretability. In the United States, insurance companies are required to be able to explain their rate and coverage decisions.{{cite news |last1=Kahn |first1=Jeremy |title=Artificial Intelligence Has Some Explaining to Do |url=https://www.bloomberg.com/news/articles/2018-12-12/artificial-intelligence-has-some-explaining-to-do |access-date=17 December 2018 |work=Bloomberg Businessweek |date=12 December 2018}} In France the Loi pour une République numérique (Digital Republic Act) grants subjects the right to request and receive information pertaining to the implementation of algorithms that process data about them.

Despite ongoing endeavors to enhance the explainability of AI models, they persist with several inherent limitations.

By making an AI system more explainable, we also reveal more of its inner workings. For example, the explainability method of feature importance identifies features or variables that are most important in determining the model's output, while the influential samples method identifies the training samples that are most influential in determining the output, given a particular input.{{cite book | last1=Bhatt | first1=Umang | last2=Xiang | first2=Alice | last3=Sharma | first3=Shubham | last4=Weller | first4=Adrian | last5=Taly | first5=Ankur | last6=Jia | first6=Yunhan | last7=Ghosh | first7=Joydeep | last8=Puri | first8=Richir | last9=M.F. Moura | first9=José | last10=Eckersley | first10=Peter | title=Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency | chapter=Explainable Machine Learning in Deployment | date=2022 | pages=648–657 | doi=10.1145/3351095.3375624 | isbn=9781450369367 | s2cid=202572724 | chapter-url=https://dl.acm.org/doi/pdf/10.1145/3351095.3375624 }} Adversarial parties could take advantage of this knowledge.

For example, competitor firms could replicate aspects of the original AI system in their own product, thus reducing competitive advantage.{{cite journal |last1=Burrel |first1=Jenna |date=2016 |title=How the machine 'thinks': Understanding opacity in machine learning algorithms |url=https://journals.sagepub.com/doi/pdf/10.1177/2053951715622512 |journal=Big Data & Society |volume=3 |issue=1 |doi=10.1177/2053951715622512 |s2cid=61330970}} An explainable AI system is also susceptible to being “gamed”—influenced in a way that undermines its intended purpose. One study gives the example of a predictive policing system; in this case, those who could potentially “game” the system are the criminals subject to the system's decisions. In this study, developers of the system discussed the issue of criminal gangs looking to illegally obtain passports, and they expressed concerns that, if given an idea of what factors might trigger an alert in the passport application process, those gangs would be able to “send guinea pigs” to test those triggers, eventually finding a loophole that would allow them to “reliably get passports from under the noses of the authorities”.{{cite book | last1=Veale | first1=Michael | last2=Van Kleek | first2=Max | last3=Binns | first3=Reuben | title=Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems | chapter=Fairness and Accountability Design Needs for Algorithmic Support in High-Stakes Public Sector Decision-Making | date=2018 | volume=40 | pages=1–14 | doi=10.1145/3173574.3174014 | isbn=9781450356206 | s2cid=3639135 | chapter-url=https://dl.acm.org/doi/pdf/10.1145/3173574.3174014 }}

Many approaches that it uses provides explanation in general, it doesn't take account for the diverse backgrounds and knowledge level of the users. This leads to challenges with accurate comprehension for all users. Expert users can find the explanations lacking in depth, and are oversimplified, while a beginner user may struggle understanding the explanations as they are complex. This limitation downplays the ability of the XAI techniques to appeal to their users with different levels of knowledge, which can impact the trust from users and who uses it. The quality of explanations can be different amongst their users as they all have different expertise levels, including different situation and conditions.{{Cite journal |last1=Yang |first1=Wenli |last2=Wei |first2=Yuchen |last3=Wei |first3=Hanyu |last4=Chen |first4=Yanyu |last5=Huang |first5=Guan |last6=Li |first6=Xiang |last7=Li |first7=Renjie |last8=Yao |first8=Naimeng |last9=Wang |first9=Xinyi |last10=Gu |first10=Xiaotong |last11=Amin |first11=Muhammad Bilal |last12=Kang |first12=Byeong |date=2023-08-10 |title=Survey on Explainable AI: From Approaches, Limitations and Applications Aspects |url=https://link.springer.com/10.1007/s44230-023-00038-y |journal=Human-Centric Intelligent Systems |language=en |volume=3 |issue=3 |pages=161–188 |doi=10.1007/s44230-023-00038-y |issn=2667-1336}}

A fundamental barrier to making AI systems explainable is the technical complexity of such systems. End users often lack the coding knowledge required to understand software of any kind. Current methods used to explain AI are mainly technical ones, geared toward machine learning engineers for debugging purposes, rather than toward the end users who are ultimately affected by the system, causing “a gap between explainability in practice and the goal of transparency”. Proposed solutions to address the issue of technical complexity include either promoting the coding education of the general public so technical explanations would be more accessible to end users, or providing explanations in layperson terms.

The solution must avoid oversimplification. It is important to strike a balance between accuracy – how faithfully the explanation reflects the process of the AI system – and explainability – how well end users understand the process. This is a difficult balance to strike, since the complexity of machine learning makes it difficult for even ML engineers to fully understand, let alone non-experts.

The goal of explainability to end users of AI systems is to increase trust in the systems, even “address concerns about lack of ‘fairness’ and discriminatory effects”. However, even with a good understanding of an AI system, end users may not necessarily trust the system.{{cite journal |last1=Hu |first1=Tongxi |last2=Zhang |first2=Xuesong |last3=Bohrer |first3=Gil |last4=Liu |first4=Yanlan |last5=Zhou |first5=Yuyu |last6=Martin |first6=Jay |last7=LI |first7=Yang |last8=Zhao |first8=Kaiguang |title=Crop yield prediction via explainable AI and interpretable machine learning: Dangers of black box models for evaluating climate change impacts on crop yield|journal=Agricultural and Forest Meteorology |date=2023 |volume=336 |page=109458 |doi=10.1016/j.agrformet.2023.109458 |s2cid=258552400 |url=https://www.sciencedirect.com/science/article/pii/S0168192323001508}} In one study, participants were presented with combinations of white-box and black-box explanations, and static and interactive explanations of AI systems. While these explanations served to increase both their self-reported and objective understanding, it had no impact on their level of trust, which remained skeptical.{{cite book | last1=Cheng | first1=Hao-Fei | last2=Wang | first2=Ruotang | last3=Zhang | first3=Zheng | last4=O’Connell | first4=Fiona | last5=Gray | first5=Terrance | last6=Harper | first6= F. Maxwell | last7=Zhu | first7=Haiyi | title=Explaining Decision-Making Algorithms through UI: Strategies to Help Non-Expert Stakeholders | series=Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems | date=2019 | volume=559 | pages=1–12 | doi=10.1145/3290605.3300789 | isbn=9781450359702 | s2cid=140281803 | url=https://dl.acm.org/doi/pdf/10.1145/3290605.3300789 }}

This outcome was especially true for decisions that impacted the end user in a significant way, such as graduate school admissions. Participants judged algorithms to be too inflexible and unforgiving in comparison to human decision-makers; instead of rigidly adhering to a set of rules, humans are able to consider exceptional cases as well as appeals to their initial decision. For such decisions, explainability will not necessarily cause end users to accept the use of decision-making algorithms. We will need to either turn to another method to increase trust and acceptance of decision-making algorithms, or question the need to rely solely on AI for such impactful decisions in the first place.

However, some emphasize that the purpose of explainability of artificial intelligence is not to merely increase users' trust in the system's decisions, but to calibrate the users' level of trust to the correct level.{{Cite journal |last1=Liel |first1=Yotam |last2=Zalmanson |first2=Lior |date=August 2023 |title=Turning Off Your Better Judgment – Conformity to Algorithmic Recommendations |url=http://journals.aom.org/doi/full/10.5465/AMPROC.2023.277bp |journal=Academy of Management Proceedings |language=en |volume=2023 |issue=1 |doi=10.5465/AMPROC.2023.277bp |s2cid=260212999 |issn=0065-0668}} According to this principle, too much or too little user trust in the AI system will harm the overall performance of the human-system unit. When the trust is excessive, the users are not critical of possible mistakes of the system and when the users do not have enough trust in the system, they will not exhaust the benefits inherent in it.

Some scholars have suggested that explainability in AI should be considered a goal secondary to AI effectiveness, and that encouraging the exclusive development of XAI may limit the functionality of AI more broadly.{{Cite journal|last1=McCoy|first1=Liam G.|last2=Brenna|first2=Connor T. A.|last3=Chen|first3=Stacy S.|last4=Vold|first4=Karina|last5=Das|first5=Sunit|date=2021-11-05|title=Believing in black boxes: machine learning for healthcare does not need explainability to be evidence-based|url=https://www.jclinepi.com/article/S0895-4356(21)00354-1/abstract|journal=Journal of Clinical Epidemiology|volume=142|language=English|issue=Online ahead of print|pages=252–257|doi=10.1016/j.jclinepi.2021.11.001|issn=0895-4356|pmid=34748907|s2cid=243810442}}{{Cite journal|last1=Ghassemi|first1=Marzyeh|last2=Oakden-Rayner|first2=Luke|last3=Beam|first3=Andrew L.|date=2021-11-01|title=The false hope of current approaches to explainable artificial intelligence in health care|url=https://www.thelancet.com/journals/landig/article/PIIS2589-7500(21)00208-9/abstract|journal=The Lancet Digital Health|language=English|volume=3|issue=11|pages=e745–e750|doi=10.1016/S2589-7500(21)00208-9|issn=2589-7500|pmid=34711379|s2cid=239963176}} Critiques of XAI rely on developed concepts of mechanistic and empiric reasoning from evidence-based medicine to suggest that AI technologies can be clinically validated even when their function cannot be understood by their operators.

Some researchers advocate the use of inherently interpretable machine learning models, rather than using post-hoc explanations in which a second model is created to explain the first. This is partly because post-hoc models increase the complexity in a decision pathway and partly because it is often unclear how faithfully a post-hoc explanation can mimic the computations of an entirely separate model. However, another view is that what is important is that the explanation accomplishes the given task at hand, and whether it is pre or post-hoc doesn't matter. If a post-hoc explanation method helps a doctor diagnose cancer better, it is of secondary importance whether it is a correct/incorrect explanation.

The goals of XAI amount to a form of lossy compression that will become less effective as AI models grow in their number of parameters. Along with other factors this leads to a theoretical limit for explainability.{{Cite journal |last=Sarkar |first=Advait |date=2022 |title=Is explainable AI a race against model complexity? |url=http://ceur-ws.org/Vol-3124/paper22.pdf |journal=Workshop on Transparency and Explanations in Smart Systems (TeXSS), in Conjunction with ACM Intelligent User Interfaces (IUI 2022) |pages=192–199 |arxiv=2205.10119 |via=CEUR Workshop Proceedings}}

Explainability was studied also in social choice theory. Social choice theory aims at finding solutions to social decision problems, that are based on well-established axioms. Ariel D. Procaccia{{Citation |last=Procaccia |first=Ariel D. |title=Axioms Should Explain Solutions |date=2019 |work=The Future of Economic Design: The Continuing Development of a Field as Envisioned by Its Researchers |series=Studies in Economic Design |pages=195–199 |editor-last=Laslier |editor-first=Jean-François |url=https://doi.org/10.1007/978-3-030-18050-8_27 |access-date=2024-03-21 |place=Cham |publisher=Springer International Publishing |language=en |doi=10.1007/978-3-030-18050-8_27 |isbn=978-3-030-18050-8 |editor2-last=Moulin |editor2-first=Hervé |editor3-last=Sanver |editor3-first=M. Remzi |editor4-last=Zwicker |editor4-first=William S.}} explains that these axioms can be used to construct convincing explanations to the solutions. This principle has been used to construct explanations in various subfields of social choice.

Cailloux and Endriss{{Cite journal |last1=Cailloux |first1=Olivier |last2=Endriss |first2=Ulle |date=2016-05-09 |title=Arguing about Voting Rules |url=https://dl.acm.org/doi/abs/10.5555/2936924.2936968 |journal=Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Systems |series=AAMAS '16 |location=Richland, SC |publisher=International Foundation for Autonomous Agents and Multiagent Systems |pages=287–295 |isbn=978-1-4503-4239-1}} present a method for explaining voting rules using the axioms that characterize them. They exemplify their method on the Borda voting rule .

Peters, Procaccia, Psomas and Zhou{{Cite journal |last1=Peters |first1=Dominik |last2=Procaccia |first2=Ariel D |last3=Psomas |first3=Alexandros |last4=Zhou |first4=Zixin |date=2020 |title=Explainable Voting |url=https://proceedings.neurips.cc/paper/2020/hash/10c72a9d42dd07a028ee910f7854da5d-Abstract.html |journal=Advances in Neural Information Processing Systems |publisher=Curran Associates, Inc. |volume=33 |pages=1525–1534}} present an algorithm for explaining the outcomes of the Borda rule using O(m²) explanations, and prove that this is tight in the worst case.

Yang, Hausladen, Peters, Pournaras, Fricker and Helbing{{Cite journal |arxiv=2310.03501 |first1=Joshua C. |last1=Yang |first2=Carina I. |last2=Hausladen |title=Designing Digital Voting Systems for Citizens: Achieving Fairness and Legitimacy in Participatory Budgeting |date=2024 |last3=Peters |first3=Dominik |last4=Pournaras |first4=Evangelos |author5=Regula Häenggli Fricker |last6=Helbing |first6=Dirk|journal=Digital Government: Research and Practice |volume=5 |issue=3 |pages=1–30 |doi=10.1145/3665332 }} present an empirical study of explainability in participatory budgeting. They compared the greedy and the equal shares rules, and three types of explanations: mechanism explanation (a general explanation of how the aggregation rule works given the voting input), individual explanation (explaining how many voters had at least one approved project, at least 10000 CHF in approved projects), and group explanation (explaining how the budget is distributed among the districts and topics). They compared the perceived trustworthiness and fairness of greedy and equal shares, before and after the explanations. They found out that, for MES, mechanism explanation yields the highest increase in perceived fairness and trustworthiness; the second-highest was Group explanation. For Greedy, Mechanism explanation increases perceived trustworthiness but not fairness, whereas Individual explanation increases both perceived fairness and trustworthiness. Group explanation decreases the perceived fairness and trustworthiness.

Nizri, Azaria and Hazon{{Cite journal |last1=Nizri |first1=Meir |last2=Hazon |first2=Noam |last3=Azaria |first3=Amos |date=2022-06-28 |title=Explainable Shapley-Based Allocation (Student Abstract) |url=https://ojs.aaai.org/index.php/AAAI/article/view/21648 |journal=Proceedings of the AAAI Conference on Artificial Intelligence |language=en |volume=36 |issue=11 |pages=13023–13024 |doi=10.1609/aaai.v36i11.21648 |s2cid=250296641 |issn=2374-3468}} present an algorithm for computing explanations for the Shapley value. Given a coalitional game, their algorithm decomposes it to sub-games, for which it is easy to generate verbal explanations based on the axioms characterizing the Shapley value. The payoff allocation for each sub-game is perceived as fair, so the Shapley-based payoff allocation for the given game should seem fair as well. An experiment with 210 human subjects shows that, with their automatically generated explanations, subjects perceive Shapley-based payoff allocation as significantly fairer than with a general standard explanation.

• {{annotated link|Algorithmic transparency}}
• {{annotated link|Right to explanation}}
• {{annotated link|Accumulated local effects}}

{{reflist}}

• {{ cite web | url=http://xaiworldconference.com/ | title=the World Conference on eXplainable Artificial Intelligence}}
• {{cite web | url=https://fatconference.org/ | title=ACM Conference on Fairness, Accountability, and Transparency (FAccT) }}
• {{ Cite journal | title= Random Forest similarity maps: A Scalable Visual Representation for Global and Local Interpretation| year= 2021| doi= 10.3390/electronics10222862| doi-access= free| last1= Mazumdar| first1= Dipankar| last2= Neto| first2= Mário Popolin| last3= Paulovich| first3= Fernando V.| journal= Electronics| volume= 10| issue= 22| page= 2862}}
• {{cite arXiv |last1=Park |first1=Dong Huk |last2=Hendricks |first2=Lisa Anne |last3=Akata |first3=Zeynep |last4=Schiele |first4=Bernt |last5=Darrell |first5=Trevor |last6=Rohrbach |first6=Marcus |title=Attentive Explanations: Justifying Decisions and Pointing to the Evidence |date=2016-12-14 |eprint=1612.04757 |class=cs.CV}}
• {{cite web | title=Explainable AI: Making machines understandable for humans | website=Explainable AI: Making machines understandable for humans | url=https://explainableai.com/ | ref={{sfnref | Explainable AI: Making machines understandable for humans}} | access-date=2017-11-02}}
• {{cite web |title=Explaining How End-to-End Deep Learning Steers a Self-Driving Car |website=Parallel Forall |date=2017-05-23 |url=https://devblogs.nvidia.com/parallelforall/explaining-deep-learning-self-driving-car/ |ref={{sfnref | Parallel Forall | 2017}} |access-date=2017-11-02}}
• {{cite web | last=Knight | first=Will | title=DARPA is funding projects that will try to open up AI's black boxes | website=MIT Technology Review | date=2017-03-14 | url=https://www.technologyreview.com/s/603795/the-us-military-wants-its-autonomous-machines-to-explain-themselves/ | access-date=2017-11-02}}
• {{cite arXiv | last1=Alvarez-Melis | first1=David | last2=Jaakkola | first2=Tommi S. |title=A causal framework for explaining the predictions of black-box sequence-to-sequence models | date=2017-07-06 | eprint=1707.01943 | class=cs.LG }}
• {{cite web | title=Similarity Cracks the Code Of Explainable AI | website=simMachines | date=2017-10-12 | url=http://simmachines.com/similarity-cracks-code-explainable-ai/ | ref={{sfnref | simMachines | 2017}} | access-date=2018-02-02}}

{{Differentiable computing}}

{{DEFAULTSORT:Explainable AI}}

Category:Quality control tools

Category:Artificial intelligence

Category:Artificial intelligence engineering