A taxonomic database is a database created to hold information on biological taxa – for example groups of organisms organized by species name or other taxonomic identifier – for efficient data management and information retrieval. Taxonomic databases are routinely used for the automated construction of biological checklists such as floras and faunas, both for print publication and online; to underpin the operation of web-based species information systems; as a part of biological collection management (for example in museums and herbaria); as well as providing, in some cases, the taxon management component of broader science or biology information systems. They are also a fundamental contribution to the discipline of biodiversity informatics. == Goals == Taxonomic databases digitize scientific biodiversity data and provide access to taxonomic data for research. Taxonomic databases vary in breadth of the groups of taxa and geographical space they seek to include, for example: beetles in a defined region, mammals globally, or all described taxa in the tree of life. A taxonomic database may incorporate organism identifiers (scientific name, author, and – for zoological taxa – year of original publication), synonyms, taxonomic opinions, literature sources or citations, illustrations or photographs, and biological attributes for each taxon (such as geographic distribution, ecology, descriptive information, threatened or vulnerable status, etc.). Some databases, such as the Global Biodiversity Information Facility(GBIF) database and the Barcode of Life Data System, store the DNA barcode of a taxon if one exists (also called the Barcode Index Number (BIN) which may be assigned, for example, by the International Barcode of Life project (iBOL) or UNITE, a database for fungal DNA barcoding). A taxonomic database aims to accurately model the characteristics of interest that are relevant to the organisms which are in scope for the intended coverage and usage of the system. For example, databases of fungi, algae, bryophytes and vascular plants ("higher plants") encode conventions from the International Code of Botanical Nomenclature while their counterparts for animals and most protists encode equivalent rules from the International Code of Zoological Nomenclature. Modelling the relevant taxonomic hierarchy for any taxon is a natural fit with the relational model employed in almost all database systems. Scientific consensus is not reached for all taxon groups, and new species continue to be described; therefore, another goal of taxonomic databases is to aid in resolving conflicts of scientific opinion and unify taxonomy. == History == Possibly the earliest documented management of taxonomic information in computerised form comprised the taxonomic coding system developed by Richard Swartz et al. at the Virginia Institute of Marine Science for the Biota of Chesapeake Bay and described in a published report in 1972. This work led directly or indirectly to other projects with greater profile including the NODC Taxonomic Code system which went through 8 versions before being discontinued in 1996, to be subsumed and transformed into the still current Integrated Taxonomic Information System (ITIS). A number of other taxonomic databases specializing in particular groups of organisms that appeared in the 1970s through to the present jointly contribute to the Species 2000 project, which since 2001 has been partnering with ITIS to produce a combined product, the Catalogue of Life. While the Catalogue of Life currently concentrates on assembling basic name information as a global species checklist, numerous other taxonomic database projects such as Fauna Europaea, the Australian Faunal Directory, and more supply rich ancillary information including descriptions, illustrations, maps, and more. Many taxonomic database projects are currently listed at the TDWG "Biodiversity Information Projects of the World" site. == Issues == The representation of taxonomic information in machine-encodable form raises a number of issues not encountered in other domains, such as variant ways to cite the same species or other taxon name, the same name used for multiple taxa (homonyms), multiple non-current names for the same taxon (synonyms), changes in name and taxon concept definition through time, and more. Non-standardized categories and metadata in taxonomic databases hampers the ability for researchers to analyze the data. One forum that has promoted discussion and possible solutions to these and related problems since 1985 is the Biodiversity Information Standards (TDWG), originally called the Taxonomic Database Working Group. While online databases have great benefits (for example, increased access to taxonomic information), they also have issues such as data integrity risks due to on- and off-line versions and continuous updates, technical access issues due to server or internet outage, and differing capacities for complex queries to extract taxonomic data into lists. As the quantity of information in online taxonomic databases rapidly expands, data aggregation, and the integration and alignment of non-standardized data across databases, is a big challenge in taxonomy and biodiversity informatics.
Weak supervision
Weak supervision (also known as semi-supervised learning) is a paradigm in machine learning, the relevance and notability of which increased with the advent of large language models due to the large amount of data required to train them. It is characterized by using a combination of a small amount of human-labeled data (exclusively used in more expensive and time-consuming supervised learning paradigm), followed by a large amount of unlabeled data (used exclusively in unsupervised learning paradigm). In other words, the desired output values are provided only for a subset of the training data. The remaining data is unlabeled or imprecisely labeled. Intuitively, it can be seen as an exam and labeled data as sample problems that the teacher solves for the class as an aid in solving another set of problems. In the transductive setting, these unsolved problems act as exam questions. In the inductive setting, they become practice problems of the sort that will make up the exam. == Problem == The acquisition of labeled data for a learning problem often requires a skilled human agent (e.g. to transcribe an audio segment) or a physical experiment (e.g. determining the 3D structure of a protein or determining whether there is oil at a particular location). The cost associated with the labeling process thus may render large, fully labeled training sets infeasible, whereas acquisition of unlabeled data is relatively inexpensive. In such situations, semi-supervised learning can be of great practical value. Semi-supervised learning is also of theoretical interest in machine learning and as a model for human learning. == Technique == More formally, semi-supervised learning assumes a set of l {\displaystyle l} independently identically distributed examples x 1 , … , x l ∈ X {\displaystyle x_{1},\dots ,x_{l}\in X} with corresponding labels y 1 , … , y l ∈ Y {\displaystyle y_{1},\dots ,y_{l}\in Y} and u {\displaystyle u} unlabeled examples x l + 1 , … , x l + u ∈ X {\displaystyle x_{l+1},\dots ,x_{l+u}\in X} are processed. Semi-supervised learning combines this information to surpass the classification performance that can be obtained either by discarding the unlabeled data and doing supervised learning or by discarding the labels and doing unsupervised learning. Semi-supervised learning may refer to either transductive learning or inductive learning. The goal of transductive learning is to infer the correct labels for the given unlabeled data x l + 1 , … , x l + u {\displaystyle x_{l+1},\dots ,x_{l+u}} only. The goal of inductive learning is to infer the correct mapping from X {\displaystyle X} to Y {\displaystyle Y} . It is unnecessary (and, according to Vapnik's principle, imprudent) to perform transductive learning by way of inferring a classification rule over the entire input space; however, in practice, algorithms formally designed for transduction or induction are often used interchangeably. == Assumptions == In order to make any use of unlabeled data, some relationship to the underlying distribution of data must exist. Semi-supervised learning algorithms make use of at least one of the following assumptions: === Continuity / smoothness assumption === Points that are close to each other are more likely to share a label. This is also generally assumed in supervised learning and yields a preference for geometrically simple decision boundaries. In the case of semi-supervised learning, the smoothness assumption additionally yields a preference for decision boundaries in low-density regions, so few points are close to each other but in different classes. === Cluster assumption === The data tend to form discrete clusters, and points in the same cluster are more likely to share a label (although data that shares a label may spread across multiple clusters). This is a special case of the smoothness assumption and gives rise to feature learning with clustering algorithms. === Manifold assumption === The data lie approximately on a manifold of much lower dimension than the input space. In this case learning the manifold using both the labeled and unlabeled data can avoid the curse of dimensionality. Then learning can proceed using distances and densities defined on the manifold. The manifold assumption is practical when high-dimensional data are generated by some process that may be hard to model directly, but which has only a few degrees of freedom. For instance, human voice is controlled by a few vocal folds, and images of various facial expressions are controlled by a few muscles. In these cases, it is better to consider distances and smoothness in the natural space of the generating problem, rather than in the space of all possible acoustic waves or images, respectively. == History == The heuristic approach of self-training (also known as self-learning or self-labeling) is historically the oldest approach to semi-supervised learning, with examples of applications starting in the 1960s. The transductive learning framework was formally introduced by Vladimir Vapnik in the 1970s. Interest in inductive learning using generative models also began in the 1970s. A probably approximately correct learning bound for semi-supervised learning of a Gaussian mixture was demonstrated by Ratsaby and Venkatesh in 1995. == Methods == === Generative models === Generative approaches to statistical learning first seek to estimate p ( x | y ) {\displaystyle p(x|y)} , the distribution of data points belonging to each class. The probability p ( y | x ) {\displaystyle p(y|x)} that a given point x {\displaystyle x} has label y {\displaystyle y} is then proportional to p ( x | y ) p ( y ) {\displaystyle p(x|y)p(y)} by Bayes' rule. Semi-supervised learning with generative models can be viewed either as an extension of supervised learning (classification plus information about p ( x ) {\displaystyle p(x)} ) or as an extension of unsupervised learning (clustering plus some labels). Generative models assume that the distributions take some particular form p ( x | y , θ ) {\displaystyle p(x|y,\theta )} parameterized by the vector θ {\displaystyle \theta } . If these assumptions are incorrect, the unlabeled data may actually decrease the accuracy of the solution relative to what would have been obtained from labeled data alone. However, if the assumptions are correct, then the unlabeled data necessarily improves performance. The unlabeled data are distributed according to a mixture of individual-class distributions. In order to learn the mixture distribution from the unlabeled data, it must be identifiable, that is, different parameters must yield different summed distributions. Gaussian mixture distributions are identifiable and commonly used for generative models. The parameterized joint distribution can be written as p ( x , y | θ ) = p ( y | θ ) p ( x | y , θ ) {\displaystyle p(x,y|\theta )=p(y|\theta )p(x|y,\theta )} by using the chain rule. Each parameter vector θ {\displaystyle \theta } is associated with a decision function f θ ( x ) = argmax y p ( y | x , θ ) {\displaystyle f_{\theta }(x)={\underset {y}{\operatorname {argmax} }}\ p(y|x,\theta )} . The parameter is then chosen based on fit to both the labeled and unlabeled data, weighted by λ {\displaystyle \lambda } : argmax Θ ( log p ( { x i , y i } i = 1 l | θ ) + λ log p ( { x i } i = l + 1 l + u | θ ) ) {\displaystyle {\underset {\Theta }{\operatorname {argmax} }}\left(\log p(\{x_{i},y_{i}\}_{i=1}^{l}|\theta )+\lambda \log p(\{x_{i}\}_{i=l+1}^{l+u}|\theta )\right)} === Low-density separation === Another major class of methods attempts to place boundaries in regions with few data points (labeled or unlabeled). One of the most commonly used algorithms is the transductive support vector machine, or TSVM (which, despite its name, may be used for inductive learning as well). Whereas support vector machines for supervised learning seek a decision boundary with maximal margin over the labeled data, the goal of TSVM is a labeling of the unlabeled data such that the decision boundary has maximal margin over all of the data. In addition to the standard hinge loss ( 1 − y f ( x ) ) + {\displaystyle (1-yf(x))_{+}} for labeled data, a loss function ( 1 − | f ( x ) | ) + {\displaystyle (1-|f(x)|)_{+}} is introduced over the unlabeled data by letting y = sign f ( x ) {\displaystyle y=\operatorname {sign} {f(x)}} . TSVM then selects f ∗ ( x ) = h ∗ ( x ) + b {\displaystyle f^{}(x)=h^{}(x)+b} from a reproducing kernel Hilbert space H {\displaystyle {\mathcal {H}}} by minimizing the regularized empirical risk: f ∗ = argmin f ( ∑ i = 1 l ( 1 − y i f ( x i ) ) + + λ 1 ‖ h ‖ H 2 + λ 2 ∑ i = l + 1 l + u ( 1 − | f ( x i ) | ) + ) {\displaystyle f^{}={\underset {f}{\operatorname {argmin} }}\left(\displaystyle \sum _{i=1}^{l}(1-y_{i}f(x_{i}))_{+}+\lambda _{1}\|h\|_{\mathcal {H}}^{2}+\lambda _{2}\sum _{i=l+1}^{l+u}(1-|f(x_{i})|)_{+}\right)} An exact solution is intractable due to the non-convex term ( 1 − | f ( x ) | ) + {\displayst
Preferential entailment
Preferential entailment is a non-monotonic logic based on selecting only models that are considered the most plausible. The plausibility of models is expressed by an ordering among models called a preference relation, hence the name preference entailment. Formally, given a propositional formula F {\displaystyle F} and an ordering over propositional models ≤ {\displaystyle \leq } , preferential entailment selects only the models of F {\displaystyle F} that are minimal according to ≤ {\displaystyle \leq } . This selection leads to a non-monotonic inference relation: F ⊨ pref G {\displaystyle F\models _{\text{pref}}G} holds if and only if all minimal models of F {\displaystyle F} according to ≤ {\displaystyle \leq } are also models of G {\displaystyle G} . Circumscription can be seen as the particular case of preferential entailment when the ordering is based on containment of the sets of variables assigned to true (in the propositional case) or containment of the extensions of predicates (in the first-order logic case).
Jan Leike
Jan Leike (born 1986 or 1987) is an AI alignment researcher who has worked at DeepMind and OpenAI. He joined Anthropic in May 2024. == Education == Jan Leike obtained his undergraduate degree from the University of Freiburg in Germany. After earning a master's degree in computer science, he pursued a PhD in machine learning at the Australian National University under the supervision of Marcus Hutter. == Career == Leike made a six-month postdoctoral fellowship at the Future of Humanity Institute before joining DeepMind to focus on empirical AI safety research, where he collaborated with Shane Legg. === OpenAI === In 2021, Leike joined OpenAI. In June 2023, he and Ilya Sutskever became the co-leaders of the newly introduced "superalignment" project, which aimed to determine how to align future artificial superintelligences within four years to ensure their safety. This project involved automating AI alignment research using relatively advanced AI systems. At the time, Sutskever was OpenAI's Chief Scientist, and Leike was the Head of Alignment. Leike was featured in Time's list of the 100 most influential personalities in AI, both in 2023 and in 2024. In May 2024, Leike announced his resignation from OpenAI, following the departure of Sutskever, Daniel Kokotajlo and several other AI safety employees from the company. Leike wrote that "Over the past years, safety culture and processes have taken a backseat to shiny products", and that he "gradually lost trust" in OpenAI's leadership. In May 2024, Leike joined Anthropic, an AI company founded by former OpenAI employees.
Alec Radford
Alec Radford is an American artificial intelligence researcher. == Biography == Radford grew up in Texas. He graduated from Cistercian Preparatory School in 2011, where he became an Eagle Scout, and dropped out of Olin College in August 2014, where he and fellow students Slater Victoroff, Diana Yuan, and Madison May had formed the startup Indico in their dorm room. In 2015, the quartet were joined by Luke Metz and the firm and the Facebook AI research lab in New York used generative adversarial networks to create realistic low pixel images. A demonstration of Indico's technology was used without proper attribution in an April 2016 demonstration by Nvidia chief executive Jensen Huang. Radford joined OpenAI around 2016, where he worked on natural-language processing. The following year, Radford trained a neural network on Amazon reviews. The model was fairly basic, with layers which allowed for human understanding. Upon exploring it, he saw that it had a special neuron linked to the sentiment of the reviews, which it had created on its own. This was a drastic improvement from previous neural networks that had analysed sentiment, because they had to be told to do so and specially trained on data that was explicitly labeled according to sentiment. This development made OpenAI chief scientist Ilya Sutskever consider that a future model, using more diverse language data, could map far more structures of meaning, eventually becoming a "learned core module" for superintelligence. In 2018, Radford was the lead author on OpenAI's seminal research paper on generative pre-trained transformers, which form the foundation of ChatGPT. At OpenAI, he worked on early GPT models, Whisper, a speech recognition model, and the image generator DALL-E. He left OpenAI in December 2024 to pursue independent research. Around March 2025, Radford joined Thinking Machines Lab as an advisor. He joined along with Bob McGrew who was previously the chief research officer of OpenAI. In April 2026, Radford, Nick Levine, and David Duvenaud released Talkie, an AI model trained on books, newspapers, scientific journals, patents, and case law published before December 31, 1930. When asked about the state of the world in 2026, it stated that one billion people would live in Europe, that London and New York would be connected by steamships that transit between the two in ten days, and "winter will be passed in Paris, and the summer in London."
Artificial intelligence content detection
Artificial intelligence detection software aims to determine whether some content (text, image, video, or audio) was generated using artificial intelligence (AI). This software is often unreliable. == Accuracy issues == Many AI detection tools have been shown to be unreliable in detecting AI-generated text. In a 2023 study conducted by Weber-Wulff et al., researchers evaluated 14 detection tools including Turnitin and GPTZero and found that "all scored below 80% of accuracy and only 5 over 70%." They also found that these tools tend to have a bias for classifying texts more as human than as AI, and that accuracy of these tools worsens upon paraphrasing. === False positives === In AI content detection, a false positive is when human-written work is incorrectly flagged as AI-written. Many AI detection platforms claim to have a minimal level of false positives, with Turnitin claiming a less than 1% false positive rate. However, later research by The Washington Post produced much higher rates of 50%, though they used a smaller sample size. False positives in an academic setting frequently lead to accusations of academic misconduct, which can have serious consequences for a student's academic record. Additionally, studies have shown evidence that many AI detection models are prone to give false positives to work written by people whose first language is not English, and also to neurodivergent people. In June 2023, Janelle Shane wrote that portions of her book You Look Like a Thing and I Love You were flagged as AI-generated. === False negatives === A false negative is a failure to identify documents with AI-written text. False negatives often happen as a result of a detection software's sensitivity level or because evasive techniques were used when generating the work to make it sound more human. False negatives are less of a concern academically, since they aren't likely to lead to accusations and ramifications. Notably, Turnitin stated they have a 15% false negative rate. == Text detection == For text, this is usually done to prevent alleged plagiarism, often by detecting repetition of words as telltale signs that a text was AI-generated (including hallucinations). Detection systems may also rely on stylistic and structural regularities associated with LLM output, such as unusually consistent grammar, formulaic transitions, repeated discourse markers, and recurring rhetorical templates. Some tools are designed less to establish authorship provenance than to flag prose that resembles common LLM-generated style patterns. They are often used by teachers marking their students, usually on an ad hoc basis. Following the release of ChatGPT and similar AI text generative software, many educational establishments have issued policies against the use of AI by students. AI text detection software is also used by those assessing job applicants, as well as online search engines, hiring, online moderation and publishing. Current detectors may sometimes be unreliable and have incorrectly marked work by humans as originating from AI while failing to detect AI-generated work in other instances. MIT Technology Review said that the technology "struggled to pick up ChatGPT-generated text that had been slightly rearranged by humans and obfuscated by a paraphrasing tool". AI text detection software has also been shown to discriminate against non-native speakers of English. Two students from the University of California, Davis, were referred to the university's Office of Student Success and Judicial Affairs (OSSJA) after their professors scanned their essays with positive results; the first with an AI detector called GPTZero, and the second with an AI detector integration in Turnitin. However, following media coverage, and a thorough investigation, the students were cleared of any wrongdoing. In April 2023, Cambridge University and other members of the Russell Group of universities in the United Kingdom opted out of Turnitin's AI text detection tool, after expressing concerns it was unreliable. The University of Texas at Austin opted out of the system six months later. In May 2023, a professor at Texas A&M University–Commerce used ChatGPT to detect whether his students' content was written by it, which ChatGPT said was the case. As such, he threatened to fail the class despite ChatGPT not being able to detect AI-generated writing. No students were prevented from graduating because of the issue, and all but one student (who admitted to using the software) were exonerated from accusations of having used ChatGPT in their content. In July 2023, a paper titled "GPT detectors are biased against non-native English writers" was released, reporting that GPTs discriminate against non-native English authors. The paper compared seven GPT detectors against essays from both non-native English speakers and essays from United States students. The essays from non-native English speakers had an average false positive rate of 61.3%. An article by Thomas Germain, published on Gizmodo in June 2024, reported job losses among freelance writers and journalists due to AI text detection software mistakenly classifying their work as AI-generated. In September 2024, Common Sense Media reported that generative AI detectors had a 20% false positive rate for Black students, compared to 10% of Latino students and 7% of White students. To improve the reliability of AI text detection, researchers have explored digital watermarking techniques. A 2023 paper titled "A Watermark for Large Language Models" presents a method to embed imperceptible watermarks into text generated by large language models (LLMs). This watermarking approach allows content to be flagged as AI-generated with a high level of accuracy, even when text is slightly paraphrased or modified. The technique is designed to be subtle and hard to detect for casual readers, thereby preserving readability, while providing a detectable signal for those employing specialized tools. However, while promising, watermarking faces challenges in remaining robust under adversarial transformations and ensuring compatibility across different LLMs. == Anti text detection == There is software available designed to bypass AI text detection. In practice, evasion may not require specialized bypass tools. Paraphrasing, style editing, and removal of repeated discourse markers can substantially reduce the effectiveness of detectors that rely on recognizable surface patterns. A study published in August 2023 analyzed 20 abstracts from papers published in the Eye Journal, which were then paraphrased using GPT-4.0. The AI-paraphrased abstracts were examined for plagiarism using QueText and for AI-generated content using Originality.AI. The texts were then re-processed through an adversarial software called Undetectable.ai in order to reduce the AI-detection scores. The study found that the AI detection tool, Originality.AI, identified text generated by GPT-4 with a mean accuracy of 91.3%. However, after reprocessing by Undetectable.ai, the detection accuracy of Originality.ai dropped to a mean accuracy of 27.8%. Some experts also believe that techniques like digital watermarking are ineffective because they can be removed or added to trigger false positives. "A Watermark for Large Language Models" paper by Kirchenbauer et al. (2023) also addresses potential vulnerabilities of watermarking techniques. The authors outline a range of adversarial tactics, including text insertion, deletion, and substitution attacks, that could be used to bypass watermark detection. These attacks vary in complexity, from simple paraphrasing to more sophisticated approaches involving tokenization and homoglyph alterations. The study highlights the challenge of maintaining watermark robustness against attackers who may employ automated paraphrasing tools or even specific language model replacements to alter text spans iteratively while retaining semantic similarity. Experimental results show that although such attacks can degrade watermark strength, they also come at the cost of text quality and increased computational resources. == Image, video, and audio detection == Several purported AI image detection software exist, to detect AI-generated images (for example, those originating from Midjourney or DALL-E). They are not completely reliable. Industry analyses have also noted that AI-driven image recognition systems often struggle in real-world environments, where inconsistent lighting, noise and variable visual inputs reduce detection reliability, a challenge highlighted in modern agricultural quality-control research. Others claim to identify video and audio deepfakes, but this technology is also not fully reliable yet either. Despite debate around the efficacy of watermarking, Google DeepMind is actively developing a detection software called SynthID, which works by inserting a digital watermark that is invisible to the human eye into the pixels of an image.
Safe and Secure Innovation for Frontier Artificial Intelligence Models Act
The Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, or SB 1047, was a failed 2024 California bill intended to "mitigate the risk of catastrophic harms from AI models so advanced that they are not yet known to exist". Specifically, the bill would have applied to models which cost more than $100 million to train and were trained using a quantity of computing power greater than 1026 integer or floating-point operations. SB 1047 would have applied to all AI companies doing business in California—the location of the company would not matter. The bill would have created protections for whistleblowers and required developers to perform risk assessments of their models prior to release, with guidance from the Government Operations Agency. It would also have established CalCompute, a University of California public cloud computing cluster for startups, researchers and community groups. == Background == The rapid increase in capabilities of AI systems in the 2020s, including the release of ChatGPT in November 2022, caused some researchers and members of the public to become concerned about the existential risks associated with increasingly powerful AI systems. Hundreds of tech executives and AI researchers, including two of the so-called "Godfathers of AI", Geoffrey Hinton and Yoshua Bengio, signed a statement in May 2023 calling for the mitigation of the "risk of extinction from AI" to be a global priority alongside "pandemics and nuclear war". However, the plausibility of these risks is still widely debated. Strong regulation of AI has been criticized for purportedly causing regulatory capture by large AI companies like OpenAI, a phenomenon in which regulation advances the interest of larger companies at the expense of smaller competition and the public in general, although OpenAI ended up opposing the bill. Other advocates of AI regulation aim to prevent bias and privacy violations, rather than existential risks. For example, some experts who view existential concerns as overblown and unrealistic view them as a distraction from near-term harms of AI like discriminatory automated decision making. In the face of existential concerns, technology companies have made voluntary commitments to conduct safety testing, for example at the AI Safety Summit and AI Seoul Summit. In 2023, not long before the bill was proposed, Governor Newsom of California and President Biden issued executive orders on artificial intelligence. State Senator Wiener said SB 1047 draws heavily on the Biden executive order, and is motivated by the absence of unified federal legislation on AI safety. Historically, California has passed regulation on several tech issues itself, including consumer privacy and net neutrality, in the absence of action by Congress. == History == === Proposal and voting === The bill was authored by State Senator Scott Wiener. Wiener first proposed AI legislation for California through an intent bill called SB 294, the Safety in Artificial Intelligence Act, in September 2023. On February 7, 2024, Wiener introduced SB 1047. On May 21, SB 1047 passed the Senate 32–1. The bill was significantly amended by Wiener on August 15, 2024, in response to industry advice. Amendments included adding clarifications, and removing the creation of a "Frontier Model Division" and the penalty of perjury. On August 28, the bill passed the State Assembly 48–16. Then, due to the amendments, the bill was once again voted on by the Senate, passing 30–9. === Veto by governor === On September 29, Governor Gavin Newsom vetoed the bill. The deadline for California lawmakers to overrule Newsom's veto was November 30, 2024. Newsom cited concerns over the bill's regulatory framework targeting only large AI models based on their computational size, while not taking into account whether the models are deployed in high-risk environments. Newsom emphasized that this approach could create a false sense of security, overlooking smaller models that might present equally significant risks. He acknowledged the need for AI safety protocols but stressed the importance of adaptability in regulation as AI technology continues to evolve rapidly. Governor Newsom also committed to working with technology experts, federal partners, and research institutions, including the Carnegie Endowment for International Peace, led by former California Supreme Court Justice Mariano-Florentino Cuéllar; and Stanford University's Human-Centered AI (HAI) Institute, led by Dr. Fei-Fei Li. He announced plans to collaborate with these entities to advance responsible AI development, aiming to protect the public while fostering innovation. == Provisions == SB 1047 would have covered AI models with training compute over 1026 integer or floating-point operations and a cost of over $100 million. If a covered model is fine-tuned using more than $10 million, the resulting model would also have been covered. The bill would have defined critical harms with respect to four categories: Creation or use of a chemical, biological, radiological, or nuclear weapon Cyberattacks on critical infrastructure causing mass casualties or at least $500 million of damage Autonomous crimes causing mass casualties or at least $500 million of damage Other harms of comparable severity Developers would have needed to create a "safety and security protocol" before training covered models. Before deployment, they would have submitted a statement of compliance, confirming they took reasonable care to take measures to prevent covered models that pose an unreasonable risk of critical harms. The statement would have included risk assessments and descriptions of their compliance process. These rules would have applied to both covered models and their derivatives, including post-training modifications, with annual third-party audits required starting in 2026. Safeguards to reduce risk included the ability to shut down the model, which has been variously described as a "kill switch" and "circuit breaker". Whistleblowing provisions would have protected employees who report safety problems and incidents. Additionally, SB 1047 would have created a public cloud computing cluster called CalCompute, associated with the University of California, to support startups, researchers, and community groups that lack large-scale computing resources. === Compliance and supervision === SB 1047 would have required developers, beginning January 1, 2026, to annually retain a third-party auditor to perform an independent audit of compliance with the requirements of the bill, as provided. The Government Operations Agency would have reviewed the results of safety tests and incidents, and issue guidance, standards, and best practices. The bill would have created a Board of Frontier Models to supervise the application of the bill by the Government Operations Agency. It is would be composed of 9 members. == Reception == === Subjects of debate === Proponents of the bill described its provisions as simple and narrowly focused, with Sen. Scott Weiner describing it as a "light-touch, basic safety bill". This was disputed by critics of the bill, who described the bill's language as vague and criticized it as consolidating power in the largest AI companies at the expense of smaller ones. Proponents, in turn, argued that the bill only applies to models trained using more than 1026 FLOPS and with over $100 million, or fine-tuned with more than $10 million, and that the threshold could be increased if needed. The penalty of perjury was also a subject of debate, and was eventually removed through an amendment. The scope of the "kill switch" requirement was also reduced, following concerns from open-source developers. The use of the term "reasonable assurance" in the bill was also controversial, and it was eventually amended to "reasonable care". Critics then argued that "reasonable care" imposed an excessive burden by requiring confidence that models could not be used to cause catastrophic harm; proponents claimed that the standard did not require certainty and that it already applied to AI developers under existing law. === Support and opposition === Individual supporters of the bill included Turing Award recipients Yoshua Bengio and Geoffrey Hinton, Elon Musk, Bill de Blasio, Kevin Esvelt, Dan Hendrycks, Vitalik Buterin, OpenAI whistleblowers Daniel Kokotajlo and William Saunders, Lawrence Lessig, Sneha Revanur, Stuart Russell, Jan Leike, actors Mark Ruffalo, Sean Astin, and Rosie Perez, Scott Aaronson, and Max Tegmark. Over 120 Hollywood celebrities, including Mark Hamill, Jane Fonda, and J. J. Abrams, also signed a statement in support of the bill. Max Tegmark likened the bill's focus on holding companies responsible for the harms caused by their models to the FDA requiring clinical trials before a company can release a drug to the market. Organizations sponsoring the bill included the Center for AI Safety, Economic Security California and Encode. The la