Explainable artificial intelligence: Difference between revisions

Revision as of 03:56, 11 August 2025

Latest revision as of 15:13, 20 August 2025

Line 3:

{{artificial intelligence}}

Within [[artificial intelligence]] (AI), '''explainable AI''' ('''XAI'''), often overlapping with '''interpretable AI''' or '''explainable machine learning''' ('''XML'''), is a field of research that explores methods that provide humans with the ability of intellectual oversight over AI algorithms.<ref>{{Cite journal|last=Longo|first=Luca|display-authors=etal|date=2024 |title=Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions|url=https://www.sciencedirect.com/science/article/pii/S1566253524000794 |journal=Information Fusion|volume=106|article-number=102301 |doi=10.1016/j.inffus.2024.102301|arxiv=2310.19775}}</ref><ref>{{Cite journal |last=Mihály |first=Héder |date=2023 |title=Explainable AI: A Brief History of the Concept |url=https://ercim-news.ercim.eu/images/stories/EN134/EN134-web.pdf |journal=ERCIM News |issue=134 |pages=9–10}}</ref> The main focus is on the [[automated reasoning|reasoning]] behind the decisions or predictions made by the AI algorithms,<ref>{{Cite journal |last1=Phillips |first1=P. Jonathon |last2=Hahn |first2=Carina A. |last3=Fontana |first3=Peter C. |last4=Yates |first4=Amy N. |last5=Greene |first5=Kristen |last6=Broniatowski |first6=David A. |last7=Przybocki |first7=Mark A. |date=2021-09-29 |title=Four Principles of Explainable Artificial Intelligence |url=https://doi.org/10.6028/NIST.IR.8312 |journal=NIST |doi=10.6028/nist.ir.8312}}</ref> to make them more understandable and transparent.<ref>{{Cite journal|last1=Vilone|first1=Giulia|last2=Longo|first2=Luca|title=Notions of explainability and evaluation approaches for explainable artificial intelligence|url=https://www.sciencedirect.com/science/article/pii/S1566253521001093|journal=Information Fusion|year=2021|volume= December 2021 - Volume 76 |pages=89–106|doi=10.1016/j.inffus.2021.05.009}}</ref> This addresses users' requirement to assess safety and scrutinize the automated decision making in applications.<ref name="auto">{{Cite journal |last1=Confalonieri |first1=Roberto |last2=Coba |first2=Ludovik |last3=Wagner |first3=Benedikt |last4=Besold |first4=Tarek R. |date=January 2021 |title=A historical perspective of explainable Artificial Intelligence |url=https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1391 |journal=WIREs Data Mining and Knowledge Discovery |language=en |volume=11 |issue=1 |article-number=e1391 |doi=10.1002/widm.1391 |issn=1942-4787|hdl=11577/3471605 |hdl-access=free }}</ref> XAI counters the "[[black box]]" tendency of machine learning, where even the AI's designers cannot explain why it arrived at a specific decision.<ref>{{Cite journal |last=Castelvecchi |first=Davide |date=2016-10-06 |title=Can we open the black box of AI? |url=http://www.nature.com/articles/538020a |journal=Nature |language=en |volume=538 |issue=7623 |pages=20–23 |doi=10.1038/538020a |pmid=27708329 |bibcode=2016Natur.538...20C |s2cid=4465871 |issn=0028-0836|doi-access=free }}</ref><ref name=guardian>{{cite news|last1=Sample|first1=Ian|title=Computer says no: why making AIs fair, accountable and transparent is crucial|url=https://www.theguardian.com/science/2017/nov/05/computer-says-no-why-making-ais-fair-accountable-and-transparent-is-crucial|access-date=30 January 2018|work=The Guardian |date=5 November 2017|language=en}}</ref>

XAI hopes to help users of AI-powered systems perform more effectively by improving their understanding of how those systems reason.<ref>{{Cite journal|last=Alizadeh|first=Fatemeh|date=2021|title=I Don't Know, Is AI Also Used in Airbags?: An Empirical Study of Folk Concepts and People's Expectations of Current and Future Artificial Intelligence|url=https://www.researchgate.net/publication/352638184|journal=Icom|volume=20 |issue=1 |pages=3–17 |doi=10.1515/icom-2021-0009|s2cid=233328352}}</ref> XAI may be an implementation of the social [[right to explanation]].<ref name=":0">{{Cite journal|last1=Edwards|first1=Lilian|last2=Veale|first2=Michael|date=2017|title=Slave to the Algorithm? Why a 'Right to an Explanation' Is Probably Not the Remedy You Are Looking For|journal=Duke Law and Technology Review|volume=16|pages=18|ssrn=2972855}}</ref> Even if there is no such legal right or regulatory requirement, XAI can improve the [[user experience]] of a product or service by helping end users trust that the AI is making good decisions.<ref>{{Cite web |last=Do Couto |first=Mark |date=February 22, 2024 |title=Entering the Age of Explainable AI |url=https://tdwi.org/Articles/2024/02/22/ADV-ALL-Entering-the-Age-of-Explainable-AI.aspx |access-date=2024-09-11 |website=TDWI}}</ref> XAI aims to explain what has been done, what is being done, and what will be done next, and to unveil which information these actions are based on.<ref name=":3">{{Cite journal|last1=Gunning|first1=D.|last2=Stefik|first2=M.|last3=Choi|first3=J.|last4=Miller|first4=T.|last5=Stumpf|first5=S.|last6=Yang|first6=G.-Z.|date=2019-12-18|title=XAI-Explainable artificial intelligence|url=https://openaccess.city.ac.uk/id/eprint/23405/|journal=Science Robotics|language=en|volume=4|issue=37|pages=eaay7120|doi=10.1126/scirobotics.aay7120|pmid=33137719|issn=2470-9476|doi-access=free}}</ref> This makes it possible to confirm existing knowledge, challenge existing knowledge, and generate new assumptions.<ref>{{Cite journal|last1=Rieg|first1=Thilo|last2=Frick|first2=Janek|last3=Baumgartl|first3=Hermann|last4=Buettner|first4=Ricardo|date=2020-12-17|title=Demonstration of the potential of white-box machine learning approaches to gain insights from cardiovascular disease electrocardiograms|journal=PLOS ONE|language=en|volume=15|issue=12|pages=e0243615|doi=10.1371/journal.pone.0243615|issn=1932-6203|pmc=7746264|pmid=33332440|bibcode=2020PLoSO..1543615R|doi-access=free}}</ref>

Line 17:

If algorithms fulfill these principles, they provide a basis for justifying decisions, tracking them and thereby verifying them, improving the algorithms, and exploring new facts.<ref>{{Cite journal|last1=Adadi|first1=A.|last2=Berrada|first2=M.|date=2018|title=Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)|journal=IEEE Access|volume=6|pages=52138–52160|doi=10.1109/ACCESS.2018.2870052|bibcode=2018IEEEA...652138A |issn=2169-3536|doi-access=free}}</ref>

Sometimes it is also possible to achieve a high-accuracy result with white-box ML algorithms. These algorithms have an interpretable structure that can be used to explain predictions.<ref name=":6">{{Cite journal|last=Rudin|first=Cynthia|date=2019|title=Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead|journal=Nature Machine Intelligence|language=en|volume=1|issue=5|pages=206–215|doi=10.1038/s42256-019-0048-x|pmid=35603010 |pmc=9122117 |arxiv=1811.10154|issn=2522-5839|doi-access=free}}</ref> Concept Bottleneck Models, which use concept-level abstractions to explain model reasoning, are examples of this and can be applied in both image<ref name="Koh Nguyen Tang Mussmann Pierson Kim Liang 2020">{{Cite conference|last1=Koh|first1=P. W.|last2=Nguyen|first2=T.|last3=Tang|first3=Y. S.|last4=Mussmann|first4=S.|last5=Pierson|first5=E.|last6=Kim|first6=B.|last7=Liang|first7=P.|date=November 2020|title=Concept bottleneck models|book-title=International Conference on Machine Learning|pages=5338–5348|publisher=PMLR}}</ref> and text<ref name="Ludan Lyu Yang Dugan Yatskar Callison-Burch 2023">{{Cite arXiv|last1=Ludan|first1=J. M.|last2=Lyu|first2=Q.|last3=Yang|first3=Y.|last4=Dugan|first4=L.|last5=Yatskar|first5=M.|last6=Callison-Burch|first6=C.|date=2023|title=Interpretable-by-Design Text Classification with Iteratively Generated Concept Bottleneck|class=cs.CL |eprint=2310.19660}}</ref> prediction tasks. This is especially important in domains like [[medicine]], [[Defense industry|defense]], [[finance]], and [[law]], where it is crucial to understand decisions and build trust in the algorithms.<ref name=":3" /> Many researchers argue that, at least for [[supervised machine learning]], the way forward is [[symbolic regression]], where the algorithm searches the space of mathematical expressions to find the model that best fits a given dataset.<ref name="Wenninger Kaymakci Wiethe 2022 p=118300">{{cite journal | last1=Wenninger | first1=Simon | last2=Kaymakci | first2=Can | last3=Wiethe | first3=Christian | title=Explainable long-term building energy consumption prediction using QLattice | journal=Applied Energy | publisher=Elsevier BV | volume=308 | year=2022 | issn=0306-2619 | doi=10.1016/j.apenergy.2021.118300 | pagearticle-number=118300| bibcode=2022ApEn..30818300W | s2cid=245428233 }}</ref><ref name="Christiansen Wilstrup Hedley 2022 p.">{{cite journal | last1=Christiansen | first1=Michael | last2=Wilstrup | first2=Casper | last3=Hedley | first3=Paula L. | title=Explainable "white-box" machine learning is the way forward in preeclampsia screening | journal=American Journal of Obstetrics and Gynecology | publisher=Elsevier BV | year=2022 | volume=227 | issue=5 | issn=0002-9378 | doi=10.1016/j.ajog.2022.06.057 | page=791| pmid=35779588 | s2cid=250160871 }}</ref><ref name="Wilstup Cave p.">{{citation | last1=Wilstup | first1=Casper | last2=Cave | first2=Chris | title=Combining symbolic regression with the Cox proportional hazards model improves prediction of heart failure deaths | publisher=Cold Spring Harbor Laboratory | date=2021-01-15 | doi=10.1101/2021.01.15.21249874 | page=| s2cid=231609904 | doi-access=free }}</ref>

AI systems optimize behavior to satisfy a mathematically specified goal system chosen by the system designers, such as the command "maximize the accuracy of [[sentiment analysis|assessing how positive]] film reviews are in the test dataset." The AI may learn useful general rules from the test set, such as "reviews containing the word "horrible" are likely to be negative." However, it may also learn inappropriate rules, such as "reviews containing '[[Daniel Day-Lewis]]' are usually positive"; such rules may be undesirable if they are likely to fail to generalize outside the training set, or if people consider the rule to be "cheating" or "unfair." A human can audit rules in an XAI to get an idea of how likely the system is to generalize to future real-world data outside the test set.<ref name="science">{{cite journal|date=5 July 2017|title=How AI detectives are cracking open the black box of deep learning|url=https://www.science.org/content/article/how-ai-detectives-are-cracking-open-black-box-deep-learning|journal=Science|language=en|access-date=30 January 2018}}.</ref>

Line 58:

=== Interpretability ===

[[File:Grokking modular addition.jpg|thumb|upright=1.2|[[Grokking (machine learning)|Grokking]] 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.<ref>{{Cite web |last=Ananthaswamy |first=Anil |date=2024-04-12 |title=How Do Machines ‘Grok’'Grok' Data? |url=https://www.quantamagazine.org/how-do-machines-grok-data-20240412/ |access-date=2025-01-21 |website=Quanta Magazine |language=en}}</ref>]]

Scholars sometimes use the term "[[mechanistic interpretability]]" to refer to the process of [[Reverse engineering|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.<ref>{{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}}</ref>

Line 80:

[[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.<ref>Minsky, et al., "The Society of Intelligent Veillance" IEEE ISTAS2013, pages 13-17.</ref>

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.<ref>{{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}}</ref> To address this issue, methods have been developed to make new models more explainable and interpretable.<ref>{{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|pagesarticle-number=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}}</ref><ref name="Lipton 31–57"/><ref name="Interpretable machine learning: def"/><ref>{{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}}</ref><ref>{{Cite arXiv |last=Abdollahi, Behnoush, and Olfa Nasraoui.|title=Explainable Restricted Boltzmann Machines for Collaborative Filtering.|eprint=1606.07129|class=stat.ML|year=2016}}</ref><ref>{{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}}</ref> 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.<ref name="Bach Binder Montavon Klauschen p=e0130140">{{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}}</ref><ref>{{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}}</ref> Other techniques explain some particular prediction made by a (nonlinear) black-box model, a goal referred to as "local interpretability".<ref>{{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}}</ref><ref>{{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}}</ref><ref>{{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|author2-link=Su-In Lee|editor2-last=Luxburg|editor2-first=U. V.|editor3-last=Bengio|editor3-first=S.|editor4-last=Wallach|editor4-first=H.|arxiv=1705.07874}}</ref><ref>{{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}}</ref><ref>{{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}}</ref><ref>{{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}}</ref> 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 <ref>{{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|url-access=subscription }}</ref> 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.<ref>{{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}}</ref><ref>{{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}}</ref>

There has been work on making glass-box models which are more transparent to inspection.<ref name=":6"/><ref>{{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|doi-access=free }}</ref> This includes [[decision tree]]s,<ref>{{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}}</ref> [[Bayesian network]]s, sparse [[linear model]]s,<ref>{{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|arxiv=1502.04269 }}</ref> and more.<ref>Bostrom, N., & Yudkowsky, E. (2014). [https://intelligence.org/files/EthicsofAI.pdf The ethics of artificial intelligence]. ''The Cambridge Handbook of Artificial Intelligence'', 316-334.</ref> The [[ACM Conference on Fairness, Accountability, and Transparency|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.<ref name="FAT* conference">{{cite web | url=https://fatconference.org/ | title=FAT* Conference }}</ref><ref>{{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}}</ref>

Some techniques allow visualisations of the inputs to which individual [[Neuron (software)|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.<ref name="Circuits">{{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 }}</ref><ref>{{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}}</ref>

There are various techniques to extract compressed representations of the features of given inputs, which can then be analysed by standard [[Cluster analysis|clustering techniques]]. Alternatively, networks can be trained to output linguistic explanations of their behaviour, which are then directly human-interpretable.<ref>{{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 }}</ref> Model behaviour can also be explained with reference to training data—for example, by evaluating which training inputs influenced a given behaviour the most,<ref>{{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}}</ref> or by approximating its predictions using the most similar instances from the training data.<ref>{{cite journal |last1=Schmaltz |first1=Allen |title=Detecting Local Insights from Global Labels: Supervised & Zero-Shot Sequence Labeling via a Convolutional Decomposition |journal=Computational Linguistics |date=23 December 2021 |volume=47 |issue=4 |pages=729-773729–773 |doi=10.1162/coli_a_00416 |url=https://doi.org/10.1162/coli_a_00416|doi-access=free }}</ref>

The use of explainable artificial intelligence (XAI) in pain research, specifically in understanding the role of electrodermal activity for [[Automated Pain Recognition|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.<ref>{{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 }}</ref>

Line 102:

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.<ref name="Explainable Machine Learning in Deployment">{{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 }}</ref> 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.<ref name="How the machine 'thinks'">{{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 |article-number=2053951715622512 |doi=10.1177/2053951715622512 |s2cid=61330970|doi-access=free }}</ref> 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”.<ref>{{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 }}</ref>

=== Adaptive integration and explanation ===

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.<ref>{{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|doi-access=free }}</ref>

=== Technical complexity ===

Line 113:

=== Understanding versus trust ===

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”.<ref name="How the machine 'thinks'" /> However, even with a good understanding of an AI system, end users may not necessarily trust the system.<ref>{{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 |pagearticle-number=109458 |doi=10.1016/j.agrformet.2023.109458 |bibcode=2023AgFM..33609458H |s2cid=258552400 |url=https://www.sciencedirect.com/science/article/pii/S0168192323001508|doi-access=free }}</ref> 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.<ref name="Explaining Decision-Making Algorithms through UI">{{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 }}</ref>

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.<ref name="Explaining Decision-Making Algorithms through UI" /> 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.<ref>{{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 |article-number=16319 |doi=10.5465/AMPROC.2023.277bp |s2cid=260212999 |issn=0065-0668|url-access=subscription }}</ref> 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.

== Criticism ==

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.<ref name=":5">{{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}}</ref><ref>{{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|doi-access=free}}</ref> 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.<ref name=":5"/>

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.<ref name=":6" /> 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.