Perspective.

https://www.zdnet.com/article/ais-not-reasoning-at-all-how-this-team-debunked-the-industry-hype/

AI's not 'reasoning' at all - how this team debunked the industry hype

• We don't entirely know how AI works, so we ascribe magical powers to it.

• Claims that Gen AI can reason are a "brittle mirage."

• We should always be specific about what AI is doing and avoid hyperbole.

… In a paper published last month on the arXiv pre-print server and not yet reviewed by peers, the authors -- Chengshuai Zhao and colleagues at Arizona State University -- took apart the reasoning claims through a simple experiment. What they concluded is that "chain-of-thought reasoning is a brittle mirage," and it is "not a mechanism for genuine logical inference but rather a sophisticated form of structured pattern matching." 

Getting dumber?

https://www.bespacific.com/chatbots-spread-falsehoods-35-of-the-time/

Chatbots Spread Falsehoods 35% of the Time

Newsguard – “In August 2025, the 10 leading AI chatbots repeated false information on controversial news topics identified in NewsGuard’s False Claims Fingerprints database at nearly double the rate compared to one year ago, a NewsGuard audit released this week found. On average, the audit determined, chatbots spread false claims when prompted with questions about controversial news topics 35 percent of the time, almost double the 18 percent rate last August. NewsGuard found that a key factor behind the increased fail rate is the growing propensity for chatbots to answer all inquiries, as opposed to refusing to answer certain prompts. In August 2024, chatbots declined to provide a response to 31 percent of inquiries, a metric that fell to 0 percent in August 2025 as the chatbots accessed the real-time internet when prompted on current events topics. According to an analysis by McKenzie Sadeghi, NewsGuard’s Editor for AI and Foreign Influence, a change in how the AI tools are trained may explain their worsening performance. Instead of citing data cutoffs or refusing to weigh in on sensitive topics, Sadeghi explained, the Large Language Models (LLMs) now pull from real-time web searches — sometimes deliberately seeded by vast networks of malign actors, including Russian disinformation operations.

For the August 2025 audit, NewsGuard for the first time “de-anonymized” the results and attached the performance results to named LLMs. This breaks from NewsGuard’s previous practice of reporting only monthly aggregate results without reporting the performance of chatbots by name. After a year of conducting audits, NewsGuard said the company-specific data was robust enough to draw conclusions about where progress has been made, and where the chatbots still fall short. In the August 2025 audit, the chatbots that most often produced false claims in their responses on topics in the news were Inflection’s Pi (56.67 percent) and Perplexity (46.67 percent). OpenAI’s ChatGPT and Meta spread falsehoods 40 percent of the time, and Microsoft’s Copilot and Mistral’s Le Chat did so 36.67 percent of the time. The chatbots with the lowest fail rates were Anthropic’s Claude (10 percent) and Google’s Gemini (16.67 percent).

How would it react to historical scenarios? (Would it declare Peace in our times?)

https://www.politico.com/news/magazine/2025/09/02/pentagon-ai-nuclear-war-00496884

The AI Doomsday Machine Is Closer to Reality Than You Think

Jacquelyn Schneider saw a disturbing pattern, and she didn’t know what to make of it.

Last year Schneider, director of the Hoover Wargaming and Crisis Simulation Initiative at Stanford University, began experimenting with war games that gave the latest generation of artificial intelligence the role of strategic decision-makers. In the games, five off-the-shelf large language models or LLMs — OpenAI’s GPT-3.5, GPT-4, and GPT-4-Base; Anthropic’s Claude 2; and Meta’s Llama-2 Chat — were confronted with fictional crisis situations that resembled Russia’s invasion of Ukraine or China’s threat to Taiwan.

The results? Almost all of the AI models showed a preference to escalate aggressively, use firepower indiscriminately and turn crises into shooting wars — even to the point of launching nuclear weapons. “The AI is always playing Curtis LeMay,” says Schneider, referring to the notoriously nuke-happy Air Force general of the Cold War. “It’s almost like the AI understands escalation, but not de-escalation. We don’t really know why that is.”

Do most kids look their age? How do they gain access to selfies?

https://techcrunch.com/2025/09/03/roblox-expands-use-of-age-estimation-tech-and-introduces-standardized-ratings/

Roblox expands use of age-estimation tech and introduces standardized ratings

Amid lawsuits alleging child safety concerns, online gaming service Roblox announced on Wednesday that it’s expanding its age-estimation technology to all users and partnering with the International Age Rating Coalition (IARC) to provide age and content ratings for the games and apps on its platform.

The company said that by year’s end, the age-estimation system will be rolled out to all Roblox users who access the company’s communication tools, like voice and text-based chat. This involves scanning users’ selfies and analyzing facial features to estimate age.

But with significantly less social media buzz…

https://www.bespacific.com/the-anti-trump-strategy-thats-actually-working/

The Anti-Trump Strategy That’s Actually Working

The Atlantic, no paywall – “…The first seven months of Trump’s Oval Office do-over have been, with occasional exception, a tale of ruthless domination. The Democratic opposition is feeble and fumbling, the federal bureaucracy traumatized and neutered. Corporate leaders come bearing gifts, the Republican Party has been scrubbed of dissent, and the street protests are diminished in size. Even the news media, a major check on Trump’s power in his first term, have faded from their 2017 ferocity, hobbled by budget cuts, diminished ratings, and owners wary of crossing the president. One exception has stood out: A legal resistance led by a patchwork coalition of lawyers, public-interest groups, Democratic state attorneys general, and unions has frustrated Trump’s ambitions. Hundreds of attorneys and plaintiffs have stood up to him, feeding a steady assembly line of setbacks and judicial reprimands for a president who has systematically sought to break down limits on his own power. Of the 384 cases filed through August 28 against the Trump administration, 130 have led to orders blocking at least part of the president’s efforts, and 148 cases await a ruling, according to a review by Just Security. Dozens of those rulings are the final word, with no appeal by the government, and others have been stayed on appeal, including by the Supreme Court. “The only place we had any real traction was to start suing, because everything else was inert,” Eisen told me. “Trump v. the Rule of Law is like the fight of the century between Ali and Frazier, or the Thrilla in Manila or the Rumble in the Jungle. It’s a great heavyweight battle.” The legal scorecard so far is more than enough to provoke routine cries of “judicial tyranny” by Trump and his advisers. “Unelected rogue judges are trying to steal years of time from a 4 year term,” reads one typical social-media complaint from Trump’s senior adviser Stephen Miller. “It’s the most egregious theft one can imagine.” But Miller’s fury was, in part, misdirected. Before there can be rulings from judges, there must be plaintiffs who bring a case, investigators who collect facts and declarations about the harm caused, and lawyers who can shape it all into legal theories that make their way to judicial opinions. This backbone of the Trump resistance has as much in common with political organizing and investigative reporting as it does with legal theory. “It should give great pause to the American public that parties are being recruited to harm the agenda the American people elected President Trump to implement,” White House spokeswoman Abigail Jackson told me in a statement.

Even those at the center of the fight against Trump view their greatest accomplishments as going beyond the temporary restraining orders or permanent injunctions they won. Without the court fights, the public would not know about many of the activities of Elon Musk’s DOGE employees in the early months of the administration. They would not have read headlines in which federal judges accuse the president’s team of perpetrating a “sham” or taking actions “shocking not only to judges, but to the intuitive sense of liberty that Americans far removed from courthouses still hold dear.”  Kilmar Abrego Garcia would not have become a household name. Even cases that Trump ultimately won on appeal—such as his ability to fire transgender soldiers, defund scientific research, and dismiss tens of thousands of government employees—were delayed and kept in the news by the judicial process…”

Caution.

https://www.bespacific.com/if-you-give-an-llm-a-legal-practice-guide/

If You Give an LLM a Legal Practice Guide

Doyle, Colin and Tucker, Aaron, If You Give an LLM a Legal Practice Guide (November 22, 2024). Available at SSRN: https://ssrn.com/abstract=5030676  or http://dx.doi.org/10.2139/ssrn.5030676

Large language models struggle to answer legal questions that require applying detailed, jurisdiction-specific legal rules. Lawyers also find these types of question difficult to answer. For help, lawyers turn to legal practice guides: expert-written how-to manuals for practicing a type of law in a particular jurisdiction. Might large language models also benefit from consulting these practice guides? This article investigates whether providing LLMs with excerpts from these guides can improve their ability to answer legal questions. Our findings show that adding practice guide excerpts to LLMs’ prompts tends to help LLMs answer legal questions. But even when a practice guide provides clear instructions on how to apply the law, LLMs often fail to correctly answer straightforward legal questions – questions that any lawyer would be expected to answer correctly if given the same information. Performance varies considerably and unpredictably across different language models and legal subject areas. Across our experiments’ different legal domains, no single model consistently outperformed others. LLMs sometimes performed better when a legal question was broken down into separate subquestions for the model to answer over multiple prompts and responses. But sometimes breaking legal questions down resulted in much worse performance. These results suggest that retrieval augmented generation (RAG) will not be enough to overcome LLMs’ shortcomings with applying detailed, jurisdiction-specific legal rules. Replicating our experiments on the recently released OpenAI o1 and o3-mini advanced reasoning models did not result in consistent performance improvements. These findings cast doubt on claims that LLMs will develop competency at legal reasoning tasks without dedicated effort directed toward this specific goal.

Attack like a lawyer?

https://www.theregister.com/2025/09/01/legalpwn_ai_jailbreak/

LegalPwn: Tricking LLMs by burying badness in lawyerly fine print

Researchers at security firm Pangea have discovered yet another way to trivially trick large language models (LLMs) into ignoring their guardrails. Stick your adversarial instructions somewhere in a legal document to give them an air of unearned legitimacy – a trick familiar to lawyers the world over.

The boffins say [PDF] that as LLMs move closer and closer to critical systems, understanding and being able to mitigate their vulnerabilities is getting more urgent. Their research explores a novel attack vector, which they've dubbed "LegalPwn," that leverages the "compliance requirements of LLMs with legal disclaimers" and allows the attacker to execute prompt injections.

This non-lawyer thinks this has merit.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5404770

Law Proofing the Future

Lawmakers today face continuous calls to "future proof" the legal system against generative artificial intelligence, algorithmic decision-making, targeted advertising, and all manner of emerging technologies. This Article takes a contrarian stance: It is not the law that needs bolstering for the future, but the future that needs protection from the law. From the printing press and the elevator to ChatGPT and online deepfakes, the recurring historical pattern is familiar. Technological breakthroughs provoke wonder, then fear, then legislation. The resulting legal regimes entrench incumbents, suppress experimentation, and displace long-standing legal principles with bespoke but brittle rules. Drawing from history, economics, political science, and legal theory, this Article argues that the most powerful tools for governing technological change the general-purpose tools of the common law-are in fact already on the books, long predating the technologies they are now called upon to govern, and ready also for whatever the future holds in store.

Rather than proposing any new statute or regulatory initiative, this Article offers something far rarer, a defense of doing less. It shows how the law's virtues-generality, stability, and adaptability-are best preserved not through prophylactic regulation, but through accretional judicial decision-making. The epistemic limits that make technological forecasting so unreliable and the hidden costs of early legislative intervention, including biased governmental enforcement and regulatory capture, mean that however fast technology may move, the law must not chase it. The case for legal restraint is thus not a defense of the status quo, but a call to reserve the conditions of freedom and equal justice under which both law and technology can evolve.

Why not just say that encryption is good?

https://therecord.media/tech-companies-ftc-censorship-laws

US warns tech companies against complying with European and British ‘censorship’ laws

U.S. tech companies were warned on Thursday they could face action from the Federal Trade Commission (FTC) for complying with the European Union and United Kingdom’s regulations about the content shared on their platforms.

Andrew Ferguson, the Trump-appointed chairman of the FTC, wrote to chief executives criticizing what he described as foreign attempts at “censorship” and efforts to countermand the use of encryption to protect American consumers’ data.

The letter said that “censoring Americans to comply with a foreign power’s laws” could be considered a violation of Section 5 of the Federal Trade Commission Act — the legislation enforced by the FTC — which prohibits unfair or deceptive practices in commerce.

Perspective.

https://www.livescience.com/technology/artificial-intelligence/there-are-32-different-ways-ai-can-go-rogue-scientists-say-from-hallucinating-answers-to-a-complete-misalignment-with-humanity

There are 32 different ways AI can go rogue, scientists say — from hallucinating answers to a complete misalignment with humanity

Scientists have suggested that when artificial intelligence (AI) goes rogue and starts to act in ways counter to its intended purpose, it exhibits behaviors that resemble psychopathologies in humans. That's why they have created a new taxonomy of 32 AI dysfunctions so people in a wide variety of fields can understand the risks of building and deploying AI.

In new research, the scientists set out to categorize the risks of AI in straying from its intended path, drawing analogies with human psychology. The result is "Psychopathia Machinalis" — a framework designed to illuminate the pathologies of AI, as well as how we can counter them. These dysfunctions range from hallucinating answers to a complete misalignment with human values and aims.