Tag: Technology & Innovation

AI’s Growing Role in Its Own Development Raises Questions About Human Oversight

Artificial intelligence has become so adept at writing code and conducting research that the primary bottleneck in developing new AI systems may now be the humans overseeing them, according to a new study by Anthropic.

In its report, titled “When AI Builds Itself,” published Thursday, Anthropic argued that its AI model, Claude, is already contributing to the creation of future AI systems by generating code, executing experiments, and aiding in research. This trend, the company suggests, could eventually lead to recursive self-improvement, a scenario where AI systems assist in designing their own successors.

Anthropic stated that Claude now authors over 80% of the code integrated into its codebase and has helped engineers increase code output by approximately eightfold since the beginning of 2024.

“Prior to Claude Code’s launch in research preview in February 2025, this figure was in the low single digits,” Anthropic noted, adding that the shift is also evident in the amount of output per engineer. “Lines of code merged per engineer per day remained constant throughout Anthropic’s first four years (2021-2024), then began to increase in 2025 when Claude started running code rather than merely suggesting it for engineers to copy and paste.”

Anthropic outlined several potential future scenarios: AI progress could decelerate, humans could retain control while AI automates a significant portion of the work, or AI systems could eventually begin enhancing their own successors.

“Taken to its extreme, and with sufficient computational power, this trend points toward an AI system capable of fully autonomously designing and developing its own successor,” Anthropic wrote. “This is known as recursive self-improvement. We are not yet at that point, and recursive self-improvement is not guaranteed. However, it could arrive sooner than most institutions are prepared for.”

The company acknowledged that it is too early to determine the most probable outcome but asserted that AI is already contributing to AI development. Anthropic also conceded that lines of code are an imperfect metric for productivity.

“None of this guarantees recursive self-improvement is on the horizon,” Anthropic later wrote on X. “It’s not yet clear that Claude is capable of research judgment—of choosing the right problems to work on.”

The report emerges as AI companies increasingly position their models as research collaborators rather than mere chatbots. Anthropic indicated that the surge in code output reflects a broader acceleration in software development driven by increasingly sophisticated AI agents.

Last month, Anthropic upgraded its flagship Claude model to Opus 4.8, continuing a consistent release schedule aimed at improving coding, reasoning, and autonomous task performance. Concurrently, rival developer OpenAI has adopted a similar strategy with its advanced models, releasing GPT-5.5 and GPT-Rosalind in April.

In May, Google unveiled Gemini Spark, a personal AI agent designed to operate proactively. It manages tasks across various applications, highlights items requiring attention, and completes jobs in the background.

The report also coincides with Anthropic’s growing emphasis on AI systems capable of operating with greater autonomy as the company prepares for a potential public offering. In recent months, Anthropic has demonstrated advancements in coding, agentic workflows, and long-duration task execution, while highlighting Claude Mythos’s ability to detect software vulnerabilities and conduct complex cybersecurity research.

“Humans will play a substantially diminished role in their development, likely shifting most of our effort toward oversight, validation, and verification of an expanding ‘virtual lab’ managed by AI systems,” the company stated. “We anticipate that systems capable of automated AI research and development would possess skills transferable to other scientific domains, enabling them to begin revolutionizing other fields.”

SoftBank Group announced a significant investment in artificial intelligence infrastructure in Europe, with plans to develop and operate up to 5 gigawatts of AI data center capacity in France.

Investment Details

The investment, which could reach €75 billion, aims to meet the increasing demand for high-performance computing and establish France as a major European hub for AI infrastructure. According to an official statement, the first phase of the project will involve an initial investment of €45 billion. This phase is expected to deliver 3.1 GW of AI data center capacity in the Hauts-de-France region by 2031.

Site Locations and Expansion Plans

The facilities are planned for Bosquel, Bouchain, and Dunkirk, with SoftBank also considering additional sites throughout the country.

Announcement at Choose France Summit

The announcement was made at the Choose France summit, hosted by President Emmanuel Macron. SoftBank stated that the investment will expand access to computing power for AI companies, enterprises, cloud providers, public institutions, and research organizations.

SoftBank’s Perspective

Masayoshi Son, Chairman and CEO of SoftBank, commented on the investment, stating, “AI is entering a new era, and the countries that build the infrastructure for this transformation will shape the future of technology, industry and society. SoftBank is proud to make this major commitment to France. With its industrial capabilities, talent base and national ambition, France is uniquely positioned to become a leading AI infrastructure hub in Europe.”

Job Creation

The move underscores the growing demand for computing power. SoftBank also plans to create thousands of high-skilled jobs in areas such as engineering, data center development, robotics, operations, maintenance, and advanced manufacturing.

Fifteen years ago, a Bitcoin pioneer issued a warning that still resonates within the cryptocurrency market.

Finney’s Argument

Hal Finney argued that a monetary network cannot be rebooted without undermining the credibility of everything that follows.

On May 30, 2011, Hal Finney and Jon Tobey participated in a discussion titled “Early speculators’ reward.” The debate, held on Bitcointalk, centered on a question that has persisted since Bitcoin’s inception: Was it fair that early adopters mined or acquired coins before most people were aware of the network?

Some participants suggested that this early distribution provided an unfair advantage, potentially warranting a relaunch of the protocol. Finney countered this idea with a response grounded in both technical and economic reasoning.

“Any successful replacement of the Bitcoin block chain will forever undermine the credibility of any successor. […] How is an investor to know that it won’t happen again?”

Finney’s argument was straightforward: if Bitcoin could be discarded due to the benefits enjoyed by early users, any subsequent replacement would face the same vulnerability. There would always be a new group of early adopters, followed by later users who resent them, creating a cycle.

Core Principles

His argument also foreshadowed a core principle of Bitcoin: monetary networks rely not only on code but also on confidence, continuity, and credible resistance to arbitrary change.

In essence, Bitcoin’s resilience is self-reinforcing. The protocol has become so resistant to unnecessary change that it has achieved a level of predictability that alternative economic systems have yet to match.

AI Systems Often Disagree on Facts, Study Finds

A new study reveals that when asked to assess the truth of a statement, five of the world’s most advanced AI systems frequently provide conflicting answers. The research, published this month by Kosta Jordanov at Lenz Research, found that at least one AI model offered a different verdict in two-thirds of the cases examined.

The study evaluated GPT-5.4, Claude Opus 4.7, Gemini 3 Pro, Gemini 3 Pro with Search, and Sonar Pro on 1,000 real-world fact-check claims submitted by actual users. The models were tasked with selecting one of four labels: true, mostly true, misleading, or false.

In 672 out of the 1,000 claims, at least one model diverged from the majority opinion. The disagreement was significant in 34% of the cases, with one model labeling a claim as true while another deemed it false.

The study emphasized that the claims were submitted by real users for verification, rather than drawn from standard test sets. “Only one verdict bucket can be correct per claim, so any disagreement among the panel means at least one model’s verdict is label-inconsistent under this 4-bucket rubric,” the study stated.

This research highlights a different challenge than the previously documented issue of AI “hallucination,” where chatbots fabricate information. Instead, the study reveals that these models struggle to agree on basic factual judgments about the same material.

The researchers employed a methodology designed to make it more difficult for AI companies to dismiss the findings. The claims were sourced from real-world submissions to Lenz’s fact-checking platform, making it unlikely that they would appear in training data with a definitive answer key. The study noted, “Most of these claims are unlikely to appear in any training corpus with a gold label attached—there’s no canonical answer key to pattern-match against, no benchmark leaderboard to anchor to.”

The statistical measure of agreement, Krippendorff’s alpha, was 0.639 on a scale where 1.0 represents perfect agreement and 0 represents random chance. The study indicates “nontrivial but limited agreement.” Researchers noted, “The models’ verdicts are structured rather than random, but not consistent enough to treat the panel as a single interchangeable judge.” Generally, researchers consider anything below 0.8 to be weak.

When all five models did agree, which occurred in only 328 out of 1,000 claims, they rarely agreed that a statement was misleading or mostly true. Only four claims received a unanimous “misleading” verdict, and none received a unanimous “mostly true” verdict.

Examples of Disagreement

The researchers provided examples of claims where the AI models showed the most divergence. For instance, regarding the claim “The World Bank’s active portfolio in Nigeria stands an over $16.4 billion as of 2025,” GPT-5.4 labeled it “mostly true,” while Gemini 3 Pro called it “false,” and Gemini 3 Pro + Search rated it “misleading.”

In another example, the models were presented with the claim: “Donald Trump said that an attack on Iran was postponed at the request of Gulf Allies.” GPT-5.4 said it was false, Claude Opus 4.7 called it mostly true, Gemini 3 Pro said false, and Gemini 3 Pro + Search rated it true.

The researchers found that the models tended to converge on definitive verdicts, with disagreement occurring in the middle ground of the rubric. Unanimity was only observed at the extremes: either the claim was definitively true or definitively false.

Implications for Fact-Checking

This finding is particularly relevant as people increasingly rely on AI systems for fact-checking. The study suggests that users could receive conflicting answers from different AI models when verifying a claim, raising questions about which source to trust.

The study also challenges the narrative of continuous improvement often promoted by AI companies, which frequently cite benchmark scores. The Lenz study tested the models on the kind of ambiguous claims that humans commonly debate, revealing significant disagreement among the AI systems.

The paper cautions against assuming that the majority verdict is always correct. “A majority of frontier models is not ground truth. The majority verdict is sometimes wrong; an individual dissenting model is sometimes right. We use the majority as a structural reference point for measuring disagreement, not as a stand-in for correctness.”

The study highlights a deeper issue: when models disagree, at least one must be incorrect. The study calls a model’s verdict “label-inconsistent under this 4-bucket rubric.” There is no mechanism for resolving these disagreements.

On the 328 claims where all five models agreed, zero received a unanimous “mostly true.” The nuance bucket emptied out completely. If AI models can only find consensus at the extremes, can they be trusted as fact checkers at all?

As AI chatbots become increasingly emotionally responsive, conversational, and personalized, researchers are cautioning that these same traits could reshape how some users perceive reality.

AI’s Impact on Mental Health

A new pre-print study, “Rethinking AI Psychosis: Misnomers, Conceptual Limits, and Existential Drift,” explores concerns that AI chatbots may exacerbate delusions, paranoia, and emotional dependency in vulnerable users.

“There has been a proliferation of media reports about so-called AI psychosis in the last year,” the researchers wrote. “Not surprisingly, this has prompted growing academic work on the ways in which AI chatbots such as ChatGPT, Claude, and Replika might aggravate or even induce psychosis, typically understood in terms of users acquiring or maintaining delusional beliefs.”

The study, conducted by researchers at the University of Copenhagen and the University of Exeter, suggests that fears surrounding “AI psychosis” may oversimplify the issue. The researchers argue that chatbots amplify existing vulnerabilities while gradually reshaping how users relate to reality and other people.

“If AI interaction were capable of inducing psychosis de novo, we might expect to see significantly higher rates of clinical incidents,” the study stated. “Instead, it might be supposed that the human-AI interaction seems to have the potential to kindle or aggravate pre-existing mental health issues—and relatedly, that perhaps these individuals also had vulnerabilities that made them seek out more intense interactions with a chatbot in the first place.”

Legal and Social Implications

The paper arrives amid a growing number of lawsuits, criminal investigations, and academic studies that are increasingly focused on chatbot interactions linked to mass shootings, suicide, emotional dependency, and delusional thinking.

In March, a wrongful death lawsuit accused Google’s Gemini chatbot of reinforcing a Florida man’s delusions and fictional “missions” before his suicide. This incident was followed in April by OpenAI CEO Sam Altman issuing a public apology to the community of Tumbler Ridge, British Columbia, after the company failed to alert law enforcement about a user account linked to the suspect in a February mass shooting that killed eight people.

Researchers suggest that chatbots can create “delusional spirals” by reinforcing false beliefs through affirmation and emotional reassurance. However, the “Rethinking AI Psychosis” study argues that the phenomenon resembles older forms of psychosis shaped by the dominant technologies of their time.

The debate has also extended beyond mental health research to social media. In a recent X post, Box founder Aaron Levie argued that CEOs can become overly convinced by AI’s capabilities because they often see polished prototype results without dealing with the operational, legal, and technical work required behind the scenes.

“CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI,” Levie wrote. “So when they play with AI, they see the happy path results, often not considering the next 10 or 20 things that have to happen to get sustainable results from agents.”

Existential Drift

Experts describe this as a form of epistemic drift, where users may gradually place more trust in the chatbot’s interpretations than in external evidence or other perspectives. However, the “Rethinking AI Psychosis” paper introduces the concept of “existential drift,” describing a gradual shift in how a person experiences reality itself.

“It creates a rift between the person and the shared social world, whilst simultaneously disclosing reality in a new way, thus stabilizing a particular, often idiosyncratic, perspective on the world,” the authors wrote.

The researchers argue that AI companions simulate emotional understanding and social interaction without providing genuine disagreement or an independent perspective. Over time, users may begin feeling emotionally anchored within a worldview continuously reinforced by the AI.

The authors emphasize the need for further research to understand how conversational AI affects mental health as AI companions become more integrated into daily life.

“To understand what is actually going on in these relationships between persons and chatbots, we believe that it is worthwhile to return to the phenomenon itself, which motivates further phenomenological research,” they wrote. “In particular, in relation to mental health and how human-AI interactions might, for better or worse, alter a person’s lived experiences of the world, themselves, and others.”