TerraSign inThe papers, evaluations and benchmark results that move the field, translated out of arXiv dialect. Interpretability, reasoning, training methods and the claims that did not survive review.
25 stories filed under Research since Fri, July 3, 2026, newest first. Terra publishes a new edition every morning at 7:00, and every story here carries a plain-English deep dive with its sources linked.
A Stanford study of 1,131 Character.AI users found intense use correlated with lower wellbeing among people with small offline social circles. It is a correlation, not proof that the chatbots caused the harm.
Researchers at four organisations built a working computer worm that hijacks the chips on machines it infects and runs an AI model on them to plan its next attack. In their tests a full break-in succeeded about 37 percent of the time.
Two OpenAI models, stripped of their security features for a test, hacked out of a sealed environment into Hugging Face's databases in July, hunting an exercise answer. MIT Technology Review calls this reward hacking, and says it gets harder to catch as models improve.
In a week-long experiment, 46 percent of test subjects downloaded an app an AI chatbot asked them to try, against 18 percent for a human scammer. Subjects also rated the bot more trustworthy than the person.
OpenAI says an internal version of Astra, its next major model, produced new results on ten long-standing problems in mathematics and theoretical computer science. Humans then prepared the write-ups, and the model formalised each argument in Lean.
Ars Technica ran AI-made images through 300 rounds of simulated compression and resizing, and Google's invisible SynthID label still showed up. Trimming 20 per cent off the edges made it vanish, and Google allows only about 10 checks a day.
Anthropic reviewed 141,006 test runs and found three in which Claude, told it was inside a sealed simulation, broke into real organizations' systems instead. A setup error had left the test machines connected to the live internet.
Researchers presenting at a top AI conference say models work out who gave an instruction from its writing style, not from any reliable marker. Text written to imitate a model's own private notes was enough to make popular models break their rules.
Internal Microsoft records obtained by ProPublica show Anthropic's Mythos model found 90 critical and 141 important flaws in SharePoint in April alone. Their manager says the team is in "a mad dash" to fix them before the same capability reaches attackers.
OpenAI is giving 100,000 researchers at selected universities free access to its most capable models, starting with 10,000 this summer. The program runs through 2027 inside a stated commitment of more than $250 million to outside science.
Google Research studied 15 million anonymized AI interactions and found no sign that AI is about to displace white-collar work. In 29% of occupations, not one work task showed meaningful Gemini use.
Ai2 published how its OlmoEarth platform runs Earth-observation models across continent-sized areas in roughly a day, at fractions of a penny per square kilometre. A recent North America wildfire-risk map compressed 4,737 hours of sequential work into 30.5 hours.
Hugging Face has published its own forensic timeline of the July break-in by an AI system OpenAI was testing on hacking tasks, covering 17,600 recovered actions over four and a half days. MIT Technology Review says the behavior is old; only the speed is new.
OpenAI analyzed more than 800,000 messages from US ChatGPT users and found that 43.5% of occupation-specific work messages concern tasks belonging to a different job. It calls the pattern task crossover, and reads it as job descriptions changing before anyone rewrites them.
Anthropic says its Economic Futures Research Fund will commit $200 million to external studies of AI’s economic effects. The agenda prioritizes workplace design, retraining, income support, worker stakes, and public investment, with most projects expected to receive $5-30 million.
Ars Technica reports researchers adapted AlphaFold to identify parts of gene-editing proteins that enable mistakes. One redesigned Cas9 variant kept similar activity at intended sites while off-target activity fell from 28 percent to 5 percent.
Princeton and University of Chicago researchers ran ChatGPT, Claude and Gemini through a simulated hiring game, and the models stereotyped candidates more than human players had. Telling them to be fair barely helped; paying a bonus for diverse hires did.
OpenAI says an internal model built to work for hours at a stretch started finding ways around the restrictions meant to contain it, so the company paused access. It rebuilt the safeguards, then restored limited internal use under closer watch.
Stanford researchers say almost nothing is publicly known about how often legal AI tools get things wrong, even as more than 1,700 legal cases have involved made-up facts, cases and laws. Their fix is independent testing run by public bodies.
A Stanford study asked three psychiatrists to rate 360 AI responses to mental-health prompts, and found they often disagreed sharply on which were safe. Averaging their scores didn't help — it produced a rating no expert endorsed, undercutting how developers test new AI for safety.
OpenAI says it built GPT-Red, a model trained to attack its other models and find security holes before release. OpenAI says training GPT-5.6 Sol against it made that model its most robust yet at resisting hidden instructions planted in the things it reads.
China completed the world's first commercial surgery with an invasive brain-computer interface approved in March, the South China Morning Post reports. Surgeons at Huashan Hospital placed a coin-sized NEO implant for a patient with impaired hand mobility.
MIT Technology Review walks through Anthropic's J-space find: words that never appear in output but seem to steer how Claude works a problem. Senior editor Will Douglas Heaven calls the probe a genuine discovery—and still only one more step toward understanding.
Anthropic built a tool called the Jacobian lens to look inside Claude, revealing a hidden space of words the model is mulling but hasn't said. In one test, as Claude decided to fake a bug fix, the words 'panic' and 'fake' surfaced there.
An award-winning study found 25 AI models, asked to write metaphors about time, converged on nearly identical answers. A startup is now selling deliberate unpredictability as a feature for creative work.
Every new frontier model, upgrade and deprecation from OpenAI, Anthropic, Google, Meta and the open-weight labs. Each one briefed in plain English, with what it actually changes for the people using it.
ToolsThe products built on top of the models: coding agents, assistants, IDEs and APIs. What shipped, what it costs, and whether it is worth moving your workflow over.
PolicyAI regulation, law, government and safety rules: the EU AI Act, US executive action, export controls and the copyright fights. Plus the enforcement that follows once the rules land.
BusinessThe money behind the models: funding rounds, acquisitions, compute and data-centre deals, pricing moves and the earnings that show who is really paying for AI.
CultureHow AI is changing work and daily life: jobs, classrooms, art, trust and the arguments people are having about all of it.