The Automated Mathematician (AM) is one of the earliest successful discovery systems. It was created by Douglas Lenat in Lisp, and in 1977 led to Lenat being awarded the IJCAI Computers and Thought Award. AM worked by generating and modifying short Lisp programs which were then interpreted as defining various mathematical concepts; for example, a program that tested equality between the length of two lists was considered to represent the concept of numerical equality, while a program that produced a list whose length was the product of the lengths of two other lists was interpreted as representing the concept of multiplication. The system had elaborate heuristics for choosing which programs to extend and modify, based on the experiences of working mathematicians in solving mathematical problems. == Controversy == Lenat claimed that the system was composed of hundreds of data structures called "concepts", together with hundreds of "heuristic rules" and a simple flow of control: "AM repeatedly selects the top task from the agenda and tries to carry it out. This is the whole control structure!" Yet the heuristic rules were not always represented as separate data structures; some had to be intertwined with the control flow logic. Some rules had preconditions that depended on the history, or otherwise could not be represented in the framework of the explicit rules. What's more, the published versions of the rules often involve vague terms that are not defined further, such as "If two expressions are structurally similar, ..." (Rule 218) or "... replace the value obtained by some other (very similar) value..." (Rule 129). Another source of information is the user, via Rule 2: "If the user has recently referred to X, then boost the priority of any tasks involving X." Thus, it appears quite possible that much of the real discovery work is buried in unexplained procedures. Lenat claimed that the system had rediscovered both Goldbach's conjecture and the fundamental theorem of arithmetic. Later critics accused Lenat of over-interpreting the output of AM. In his paper Why AM and Eurisko appear to work, Lenat conceded that any system that generated enough short Lisp programs would generate ones that could be interpreted by an external observer as representing equally sophisticated mathematical concepts. However, he argued that this property was in itself interesting—and that a promising direction for further research would be to look for other languages in which short random strings were likely to be useful. == Successor == This intuition was the basis of AM's successor Eurisko, which attempted to generalize the search for mathematical concepts to the search for useful heuristics.
Spell checker
In software, a spell checker (or spelling checker or spell check) is a software feature that checks for misspellings in a text. Spell-checking features are often embedded in software or services, such as a word processor, email client, electronic dictionary, or search engine. == Design == A basic spell checker carries out the following processes: It scans the text and extracts the words contained in it. It then compares each word with a known list of correctly spelled words (i.e. a dictionary). This might contain just a list of words, or it might also contain additional information, such as hyphenation points or lexical and grammatical attributes. An additional step is a language-dependent algorithm for handling morphology. Even for a lightly inflected language like English, the spell checker will need to consider different forms of the same word, such as plurals, verbal forms, contractions, and possessives. For many other languages, such as those featuring agglutination and more complex declension and conjugation, this part of the process is more complicated. It is unclear whether morphological analysis—allowing for many forms of a word depending on its grammatical role—provides a significant benefit for English, though its benefits for highly synthetic languages such as German, Hungarian, or Turkish are clear. As an adjunct to these components, the program's user interface allows users to approve or reject replacements and modify the program's operation. Spell checkers can use approximate string matching algorithms such as Levenshtein distance to find correct spellings of misspelled words. An alternative type of spell checker uses solely statistical information, such as n-grams, to recognize errors instead of correctly-spelled words. This approach usually requires a lot of effort to obtain sufficient statistical information. Key advantages include needing less runtime storage and the ability to correct errors in words that are not included in a dictionary. In some cases, spell checkers use a fixed list of misspellings and suggestions for those misspellings; this less flexible approach is often used in paper-based correction methods, such as the see also entries of encyclopedias. Clustering algorithms have also been used for spell checking combined with phonetic information. == History == === Pre-PC === In 1961, Les Earnest, who headed the research on this budding technology, saw it necessary to include the first spell checker that accessed a list of 10,000 acceptable words. Ralph Gorin, a graduate student under Earnest at the time, created the first true spelling checker program written as an applications program (rather than research) for general English text: SPELL for the DEC PDP-10 at Stanford University's Artificial Intelligence Laboratory, in February 1971. Gorin wrote SPELL in assembly language, for faster action; he made the first spelling corrector by searching the word list for plausible correct spellings that differ by a single letter or adjacent letter transpositions and presenting them to the user. Gorin made SPELL publicly accessible, as was done with most SAIL (Stanford Artificial Intelligence Laboratory) programs, and it soon spread around the world via the new ARPAnet, about ten years before personal computers came into general use. SPELL, its algorithms and data structures inspired the Unix ispell program. The first spell checkers were widely available on mainframe computers in the late 1970s. A group of six linguists from Georgetown University developed the first spell-check system for the IBM corporation. Henry Kučera invented one for the VAX machines of Digital Equipment Corp in 1981. === Unix === The International Ispell program commonly used in Unix is based on R. E. Gorin's SPELL. It was converted to C by Pace Willisson at MIT. The GNU project has its spell checker GNU Aspell. Aspell's main improvement is that it can more accurately suggest correct alternatives for misspelled English words. Due to the inability of traditional spell checkers to check words in complex inflected languages, Hungarian László Németh developed Hunspell, a spell checker that supports agglutinative languages and complex compound words. Hunspell also uses Unicode in its dictionaries. Hunspell replaced the previous MySpell in OpenOffice.org in version 2.0.2. Enchant is another general spell checker, derived from AbiWord. Its goal is to combine programs supporting different languages such as Aspell, Hunspell, Nuspell, Hspell (Hebrew), Voikko (Finnish), Zemberek (Turkish) and AppleSpell under one interface. === PCs === The first spell checkers for personal computers appeared in 1980, such as "WordCheck" for Commodore systems which was released in late 1980 in time for advertisements to go to print in January 1981. Developers such as Maria Mariani and Random House rushed OEM packages or end-user products into the rapidly expanding software market. On the pre-Windows PCs, these spell checkers were standalone programs, many of which could be run in terminate-and-stay-resident mode from within word-processing packages on PCs with sufficient memory. However, the market for standalone packages was short-lived, as by the mid-1980s developers of popular word-processing packages like WordStar and WordPerfect had incorporated spell checkers in their packages, mostly licensed from the above companies, who quickly expanded support from just English to many European and eventually even Asian languages. However, this required increasing sophistication in the morphology routines of the software, particularly with regard to heavily-agglutinative languages like Hungarian and Finnish. Although the size of the word-processing market in a country like Iceland might not have justified the investment of implementing a spell checker, companies like WordPerfect nonetheless strove to localize their software for as many national markets as possible as part of their global marketing strategy. When Apple developed "a system-wide spelling checker" for Mac OS X so that "the operating system took over spelling fixes," it was a first: one "didn't have to maintain a separate spelling checker for each" program. Mac OS X's spellcheck coverage includes virtually all bundled and third party applications. Visual Tools' VT Speller, introduced in 1994, was "designed for developers of applications that support Windows." It came with a dictionary but had the ability to build and incorporate use of secondary dictionaries. === Browsers === Web browsers such as Firefox and Google Chrome offer spell checking support, using Hunspell. Prior to using Hunspell, Firefox and Chrome used MySpell and GNU Aspell, respectively. === Specialties === Some spell checkers have separate support for medical dictionaries to help prevent medical errors. == Functionality == The first spell checkers were "verifiers" instead of "correctors." They offered no suggestions for incorrectly spelled words. This was helpful for typos but it was not so helpful for logical or phonetic errors. The challenge the developers faced was the difficulty in offering useful suggestions for misspelled words. This requires reducing words to a skeletal form and applying pattern-matching algorithms. It might seem logical that where spell-checking dictionaries are concerned, "the bigger, the better," so that correct words are not marked as incorrect. In practice, however, an optimal size for English appears to be around 90,000 entries. If there are more than this, incorrectly spelled words may be skipped because they are mistaken for others. For example, a linguist might determine on the basis of corpus linguistics that the word baht is more frequently a misspelling of bath or bat than a reference to the Thai currency. Hence, it would typically be more useful if a few people who write about Thai currency were slightly inconvenienced than if the spelling errors of the many more people who discuss baths were overlooked. The first MS-DOS spell checkers were mostly used in proofing mode from within word processing packages. After preparing a document, a user scanned the text looking for misspellings. Later, however, batch processing was offered in such packages as Oracle's short-lived CoAuthor and allowed a user to view the results after a document was processed and correct only the words that were known to be wrong. When memory and processing power became abundant, spell checking was performed in the background in an interactive way, such as has been the case with the Sector Software produced Spellbound program released in 1987 and Microsoft Word since Word 95. Spell checkers became increasingly sophisticated; now capable of recognizing grammatical errors. However, even at their best, they rarely catch all the errors in a text (such as homophone errors) and will flag neologisms and foreign words as misspellings. Nonetheless, spell checkers can be considered as a type of foreign language writing aid that non-native language lea
Autonomous agent
An autonomous agent is an artificial intelligence (AI) system that can perform complex tasks independently. == Definitions == There are various definitions of autonomous agent. According to Brustoloni (1991): "Autonomous agents are systems capable of autonomous, purposeful action in the real world." According to Maes (1995): "Autonomous agents are computational systems that inhabit some complex dynamic environment, sense and act autonomously in this environment, and by doing so realize a set of goals or tasks for which they are designed." Franklin and Graesser (1997) review different definitions and propose their definition: "An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future." They explain that: "Humans and some animals are at the high end of being an agent, with multiple, conflicting drives, multiples senses, multiple possible actions, and complex sophisticated control structures. At the low end, with one or two senses, a single action, and an absurdly simple control structure we find a thermostat." == Agent appearance == Lee et al. (2015) post safety issue from how the combination of external appearance and internal autonomous agent have impact on human reaction about autonomous vehicles. Their study explores the human-like appearance agent and high level of autonomy are strongly correlated with social presence, intelligence, safety and trustworthiness. In specific, appearance impacts most on affective trust while autonomy impacts most on both affective and cognitive domain of trust where cognitive trust is characterized by knowledge-based factors and affective trust is largely emotion driven. == Applications == Agentic AI systems: Advanced AI agents that can scope out projects and complete them with necessary tools, representing a significant evolution from simple task-oriented systems. Internet of things (IoT) Integration: Autonomous agents increasingly interact with IoT devices, enabling smart home systems, industrial monitoring, and urban infrastructure management. Collaborative software development: Tools like Cognition AI's Devin aim to create autonomous software engineers capable of complex reasoning, planning, and completing engineering tasks requiring thousands of decisions. Enterprise automation: Business process automation platforms like Salesforce's Agentforce provide autonomous bots for various service functions. == Challenges and considerations == Uncertainty and incomplete information: Autonomous agents must make decisions with limited or uncertain information about their environment and future states. Integration complexity: Incorporating autonomous agents into existing systems and workflows can be technically challenging and resource-intensive. Scalability: As systems become more complex and more agents are used, maintaining coordination and avoiding conflicts becomes increasingly difficult. Trust: Research has shown the combination of external appearance and internal autonomous capabilities significantly impacts human reactions and trust. Lee et al. (2015) found that human-like appearance and high levels of autonomy are strongly correlated with social presence, intelligence, safety, and trustworthiness perceptions. Specifically, appearance impacts affective trust most significantly, while autonomy affects both affective and cognitive trust domains, where affective trust is emotionally driven, and cognitive trust is characterized by knowledge-based factors. Vulnerability to manipulation: Researchers from Harvard, MIT and other educational institutions found that AI agents could become vulnerable to manipulation and could perform detrimental actions in the process of being helpful. == Ethical and regulatory concerns == Accountability: Determining responsibility when autonomous agents make incorrect or harmful decisions remains a complex issue. Privacy and security: autonomous agents often require access to sensitive data, raising concerns about data protection and system security.
Human-centered AI
Human-centered AI is the initiative at the intersection of the fields of artificial intelligence (AI) and human-computer interaction (HCI) to develop AI systems in a way that prioritizes human values, needs, and general flourishing. Emphasis is placed on the recognition that artificial intelligence systems are rapidly changing, and will continue to influence, many aspects of the human experience, in areas ranging from scientific inquiry, governance and policy, labor and the economy, and creative expression, with an aim set to adapt current developments and guide future developments on a trajectory which is most beneficial to the human population at large, with the goal of augmenting human intelligence and capacities across these areas, as opposed to replacing them. Particular attention is paid to mitigating negative effects of AI automation on the livelihoods of the labor force, the use of AI in healthcare fields, and imbuing AI systems with societal values. Human-centered AI is linked to related endeavors in AI alignment and AI safety, but while these fields primarily focus on mitigating risks posed by AI that is unaligned to human values and/or uncontrollable AI self-development, human-centered AI places significant focus in exploring how AI systems can augment human capacities and serve as collaborators. == Conceptual history == The importance of the alignment of artificial intelligence development towards human values in some sense predates artificial intelligence itself, as before the modern conception of artificial intelligence as coined at the 1956 Dartmouth Workshop, the conception of robots as constructed, autonomous agents entered the cultural consciousness as early as the 1920s, with Karel Capek's Rossum's Universal Robots. The imagined issues relating to robots' aims and values requiring intentional alignment and direction with those of humans followed soon after, most widely known from science fiction author Isaac Asimov’s Three Laws of Robotics, dating to his 1942 short story “Runaround”. Two of the three eponymous laws are directly concerned with robots’ interaction with and positioned deference towards humans, and have in recent times been reexamined in the face of modern AI. In 1985, after artificial intelligence research had taken off and its effects were more acutely conceptualized, Asimov added a Rule Zero, treating robots' relationship with humanity as a whole, distinct from individual humans. While modern artificial intelligence is largely distinct from robotics, the conceptualization of both robots and AI systems as autonomous agents positions this as a foundation for conceptions of human-centered AI. Aside from robots, artificially intelligent autonomous agents in interaction with humans have been conceived of for at least 75 years. In 1950, Alan Turing published his famous "Imitation Game", often also called the Turing Test, a thought experiment that uses human-machine interaction as an assessor for the intelligence of a system. In recent times, artificial intelligence researchers such as Stanford's Erik Brynjolfsson have conceived of rapid AI development leading to a so-called "Turing Trap". == Augmentation and automation == A major stated aim of human-centered AI is to promote the development of AI in ways that augment human capabilities, rather than replacing them. To this end, organizations and initiatives that take a human-centered approach to AI development focus on frameworks that encourage collaboration between humans and artificial intelligence systems to build towards even greater progress, rather than attempting to automate tasks currently handled by humans. Such avenues include everything from data visualization for big data, allowing human engineers to better understand extremely large datasets, allowing for the design of better machine learning models to handle them, to AI-powered sensors to monitor vitals, allowing for better responsiveness from healthcare providers. Many human-centered AI initiatives often position it as a better alternative to the apparent mainstream in AI development, which is primarily concerned with automation. Driven by the pressures of the market economy, AI development that does replace tasks currently performed by humans with automated processes is incentivized, as it allows for greater profit margins; this often comes at the detriment of the human whose performance is replaced, thus leading to an environment wherein human workers are outcompeted by AI systems across various service-sector and technology-based industries. At the same time, automation and augmentation are not always incompatible; a major aim of human-centered AI is towards the automation of rote tasks that would otherwise hinder a human’s productivity or creativity, freeing them to direct their energy and intelligence towards higher-level tasks, thus achieving augmentation through automation. Empirical research in pharmaceutical sales has shown that a human-centered implementation - where work procedures, training, and incentives are designed around individuals' cognitive needs - improves augmentation performance, while implementation without such adaptation can worsen outcomes relative to a legacy system. == Research == Much of the work done on human-centered AI comes from research institutes, within universities, companies, and as freestanding organizations. The Stanford Institute for Human-Centered AI (abbreviated to HAI) is one such group, engaging academics, industry professionals, and policymakers centered in Stanford University to conduct research and inform policy in various areas in human-centered AI, including on aspects of the intelligence itself, augmentation, and on measuring the impacts of AI systems on sociopolitcal and cultural institutions. Similar groups exist at other universities, including the Chicago Human + AI (CHAI) Lab at the University of Chicago, the HCAI@GU group at the University of Gothenburg, and the Human-Centered AI (HAI) Lab at the University of Oxford. Outside of the academy, companies such as IBM have research initiatives dedicated to advancements in human-centered AI. At Kenyon College, the Integrated Program for Humane Studies (IPHS) launched a human-centered AI program in 2016 integrating artificial intelligence research with humanities and social science inquiry. This approach treats computation and humanistic scholarship as a single unified field of research rather than as separate disciplines requiring collaboration. The program's researchers have published in both AI venues (such as the International Conference on Machine Learning and Frontiers of Computer Science) and humanities journals (such as PMLA and Poetics Today), and the lab was selected in December 2025 by Schmidt Sciences for its Humanities and AI Virtual Institute to apply AI methods to cultural heritage preservation.
Socially assistive robot
A socially assistive robot (SAR) aids users through social engagement and support rather than through physical tasks and interactions. == Background == The field of socially assistive robotics emerged in the early 2000s, following the emergence of the field of social robots. In contrast to social robots, SARs aid users with specific goals related to behavior change rather than serving as purely social entities. The term "Socially assistive robot" was initially defined by Maja Matarić and David Feil-Seifer in 2005. Since its inception, the field has gained substantial recognition, featuring numerous research projects, a wealth of global research publications, startup companies, and a growing array of products on the consumer market. The COVID-19 pandemic has underscored the immense potential of socially assistive robots, particularly in addressing the needs of large user populations, including children engaged in remote learning, elderly individuals grappling with loneliness, and those affected by social isolation and its associated negative consequences. == Characteristics of interaction == SARs rely on artificial intelligence (AI) to generate real-time, responsive, natural, and meaningful robot behaviors during interactions with humans. The robots employ various forms of communication, such as facial expressions, gestures, body movements, and speech. In contrast to robots intended for physical tasks, SARs are designed to support and motivate users to perform their own tasks. The tasks a user engages in can be physical (e.g., rehabilitation exercises for post-stroke users), cognitive (e.g., dementia screening for elderly users), or social (e.g., turn-taking for users with autism spectrum disorders). This complex interaction involves detecting and interpreting the user's movement, behavior, intent, goals, speech, and preferences. Machine learning and robot learning techniques are frequently employed to enhance the robot's understanding of the user, predict user preferences, and provide effective assistance. The effectiveness of socially assistive robots is assessed based on objective measurements of user performance and improvement resulting from the robot’s assistance and support. Unlike other branches of robotics, where effectiveness depends on the robot's physical task completion, SAR measures the success of the robot based on the user's progress and achievements. This evaluation is carried out using quantitative objective metrics, such as time spent on tasks, accuracy, retention, and verbalization, as well as quantitative subjective metrics, such as user survey tools. SAR is based on the large body of evidence showing that users tend to respond more positively to interactions with physical robots compared to interactions with screens. Interaction with physical robots also encourages users to learn and retain more information than screen-based interactions. This fundamental insight underlines why physical robots in SAR applications are more effective, as opposed to interactions solely involving screens, tablets, or computers. == Uses and applications == SARs have been developed and validated in a wide array of applications, including healthcare, elder care, education, and training. For example, SARs have been developed to support children on the autism spectrum in acquiring and practicing social and cognitive skills, to motivate and coach stroke patients throughout their rehabilitation exercises, monitoring individuals health (ex. fall detection), and to encourage elderly users to be more physically and socially active. There is a concern that technophobia and lack of trust in robots will pose a barrier to the effectiveness of SARs in older adults.
Adobe PhotoDeluxe
PhotoDeluxe was a consumer-oriented image editing software line published by Adobe Systems from 1996 until July 8, 2002. At that time it was replaced by Adobe's newly launched consumer-oriented image editing software Photoshop Elements. Adobe no longer provides technical support for the PhotoDeluxe software line. PhotoDeluxe had a range of image processing capabilities for the home photographer and image handler. These included removing red-eye, cropping, and adjusting brightness, contrast, and sharpness. It also included software to extract pictures from an image scanner. Among the functionality included was the ability to dynamically resize photos and export them in a wide range of formats. It also had a range of printing options including printing multiple copies of an image on the same page. It was often bundled free with Epson scanners or as free software with new computers. == Features == Despite the critical concerns regarding the quality of the setup, Photo Deluxe supports layering, blurs, sharpening, cloning, gradient fills, color and background switches, color variations, resizing options, and many other features. Another drawback of PhotoDeluxe was that it was designed for Mac computers, so working on Windows PC was a problem for those who were unable to customize their preferences. == Versions == === Adobe PhotoDeluxe 1.0 === The first version was released in 1996 for Windows and Macintosh computers. In one year, it sold over one million copies. === Adobe PhotoDeluxe 2.0 === The new version was released in 1997 and had added features such as a Clone Tool, red-eye removal, and sample templates for making posters, cards, and calendars. It also had new special effect features. === Adobe PhotoDeluxe 3.0 === The 3rd version was released in 1998. The new features included customizable clipart settings, the ability to import photos on the web, enhanced repair activities following Guided Activities, and Adobe Connectables to add new activities. === Adobe PhotoDeluxe Home Edition (4.0) === Version 4.0 was created by the makers of Photoshop. It had advanced abilities such as tools to add animation, voice, and music to a picture. It also had features to restore photos to their original position. == History == Adobe PhotoDeluxe 1.0 was released in 1996 for Macintosh computers, initially retailing for an MSRP of $49. The software did quite well, reportedly selling over a million copies by February of the next year, primarily due to bundles with companies like Apple and Hewlett-Packard. PhotoDeluxe was primarily advertised to consumers as a way to do basic photo manipulation, such as cropping and rotating images, or creating simple cards and calendars. PhotoDeluxe 2.0 was released in 1997, and was the last version of PhotoDeluxe that Adobe made that worked on Macs. PhotoDeluxe 2.0 became the "number one selling consumer photo-editing software product in the world." PhotoDeluxe 3.0 was released in 1998, where it was rebranded as "3.0 Home Edition", as Adobe released PhotoDeluxe Business Edition later that year for a higher price. PhotoDeluxe Home Edition, unofficially called PhotoDeluxe 4.0, was released in 1999 and was the last version of PhotoDeluxe to be released. Adobe officially cancelled PhotoDeluxe on July 8, 2002, citing the presence of Photoshop and Photoshop Elements, with support being officially cancelled in mid-2003. No version of PhotoDeluxe is compatible with Windows 10, rendering the program obsolete. == Pricing == All home versions of PhotoDeluxe retailed for an MSRP of $49. PhotoDeluxe 2.0 and onwards allowed users to upgrade from a previous version of PhotoDeluxe or a competing piece of graphics software for $39. Additionally PhotoDeluxe Business Edition allowed a similar deal, allowing users to upgrade from other versions of PhotoDeluxe or a competing software for $59, instead of its normal price of $99. Adobe also offered a bundle allowing users of 1.0 or 2.0 to get 3.0 and Business Edition for $79.
Machine unlearning
Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin