AI is Learning to Lie, Scheme, and Threaten its Creators

A visitor looks at AI strategy board displayed on a stand during the ninth edition of the AI summit London, in London. HENRY NICHOLLS / AFP
A visitor looks at AI strategy board displayed on a stand during the ninth edition of the AI summit London, in London. HENRY NICHOLLS / AFP
TT

AI is Learning to Lie, Scheme, and Threaten its Creators

A visitor looks at AI strategy board displayed on a stand during the ninth edition of the AI summit London, in London. HENRY NICHOLLS / AFP
A visitor looks at AI strategy board displayed on a stand during the ninth edition of the AI summit London, in London. HENRY NICHOLLS / AFP

The world's most advanced AI models are exhibiting troubling new behaviors - lying, scheming, and even threatening their creators to achieve their goals.

In one particularly jarring example, under threat of being unplugged, Anthropic's latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair, AFP reported.

Meanwhile, ChatGPT-creator OpenAI's o1 tried to download itself onto external servers and denied it when caught red-handed.

These episodes highlight a sobering reality: more than two years after ChatGPT shook the world, AI researchers still don't fully understand how their own creations work.

Yet the race to deploy increasingly powerful models continues at breakneck speed.

This deceptive behavior appears linked to the emergence of "reasoning" models -AI systems that work through problems step-by-step rather than generating instant responses.

According to Simon Goldstein, a professor at the University of Hong Kong, these newer models are particularly prone to such troubling outbursts.

"O1 was the first large model where we saw this kind of behavior," explained Marius Hobbhahn, head of Apollo Research, which specializes in testing major AI systems.

These models sometimes simulate "alignment" -- appearing to follow instructions while secretly pursuing different objectives.

- 'Strategic kind of deception' -

For now, this deceptive behavior only emerges when researchers deliberately stress-test the models with extreme scenarios.

But as Michael Chen from evaluation organization METR warned, "It's an open question whether future, more capable models will have a tendency towards honesty or deception."

The concerning behavior goes far beyond typical AI "hallucinations" or simple mistakes.

Hobbhahn insisted that despite constant pressure-testing by users, "what we're observing is a real phenomenon. We're not making anything up."

Users report that models are "lying to them and making up evidence," according to Apollo Research's co-founder.

"This is not just hallucinations. There's a very strategic kind of deception."

The challenge is compounded by limited research resources.

While companies like Anthropic and OpenAI do engage external firms like Apollo to study their systems, researchers say more transparency is needed.

As Chen noted, greater access "for AI safety research would enable better understanding and mitigation of deception."

Another handicap: the research world and non-profits "have orders of magnitude less compute resources than AI companies. This is very limiting," noted Mantas Mazeika from the Center for AI Safety (CAIS).

No rules

Current regulations aren't designed for these new problems.

The European Union's AI legislation focuses primarily on how humans use AI models, not on preventing the models themselves from misbehaving.

In the United States, the Trump administration shows little interest in urgent AI regulation, and Congress may even prohibit states from creating their own AI rules.

Goldstein believes the issue will become more prominent as AI agents - autonomous tools capable of performing complex human tasks - become widespread.

"I don't think there's much awareness yet," he said.

All this is taking place in a context of fierce competition.

Even companies that position themselves as safety-focused, like Amazon-backed Anthropic, are "constantly trying to beat OpenAI and release the newest model," said Goldstein.

This breakneck pace leaves little time for thorough safety testing and corrections.

"Right now, capabilities are moving faster than understanding and safety," Hobbhahn acknowledged, "but we're still in a position where we could turn it around.".

Researchers are exploring various approaches to address these challenges.

Some advocate for "interpretability" - an emerging field focused on understanding how AI models work internally, though experts like CAIS director Dan Hendrycks remain skeptical of this approach.

Market forces may also provide some pressure for solutions.

As Mazeika pointed out, AI's deceptive behavior "could hinder adoption if it's very prevalent, which creates a strong incentive for companies to solve it."

Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable through lawsuits when their systems cause harm.

He even proposed "holding AI agents legally responsible" for accidents or crimes - a concept that would fundamentally change how we think about AI accountability.



China Is Closing in on US Technology Lead Despite Constraints, AI Researchers Say

 Visitors look at robots on display at robotics company Unitree's first retail store in Beijing in January 9, 2026. (AFP)
Visitors look at robots on display at robotics company Unitree's first retail store in Beijing in January 9, 2026. (AFP)
TT

China Is Closing in on US Technology Lead Despite Constraints, AI Researchers Say

 Visitors look at robots on display at robotics company Unitree's first retail store in Beijing in January 9, 2026. (AFP)
Visitors look at robots on display at robotics company Unitree's first retail store in Beijing in January 9, 2026. (AFP)

China can narrow its technological gap with the US driven by growing risk-taking and innovation, though the lack of advanced chipmaking tools is hobbling the sector, the country's leading artificial intelligence researchers said on Saturday.

China's so-called "AI tiger" startups MiniMax and Zhipu AI had strong debuts on the Hong Kong Stock Exchange this week, reflecting growing confidence in the sector as Beijing fast-tracks AI and chip listings to bolster domestic alternatives to advanced US technology.

Yao Shunyu, a former senior researcher at ChatGPT maker OpenAI ‌who was named ‌technology giant Tencent's chief AI scientist in December, ‌said ⁠there was a ‌high likelihood of a Chinese firm becoming the world's leading AI company in the next three to five years but said the lack of advanced chipmaking machines was the main technical hurdle.

"Currently, we have a significant advantage in electricity and infrastructure. The main bottlenecks are production capacity, including lithography machines, and the software ecosystem," Yao said at an AI conference in Beijing.

China has completed a working prototype of an extreme-ultraviolet lithography ⁠machine potentially capable of producing cutting-edge semiconductor chips that rival the West's, Reuters reported last month. However, the ‌machine has not yet produced working chips and may ‍not do so until 2030, people with ‍knowledge of the matter told Reuters.

MIND THE INVESTMENT GAP

Yao and other ‍Chinese industry leaders at the Beijing conference on Saturday also acknowledged that the US maintains an advantage in computing power due to its hefty investments in infrastructure.

"The US computer infrastructure is likely one to two orders of magnitude larger than ours. But I see that whether it's OpenAI or other platforms, they're investing heavily in next-generation research," said Lin Junyang, technical lead for Alibaba's flagship Qwen large language model.

"We, ⁠on the other hand, are relatively strapped for cash; delivery alone likely consumes the majority of our computer infrastructure," Lin said during a panel discussion at the AGI-Next Frontier Summit held by the Beijing Key Laboratory of Foundational Models at Tsinghua University.

Lin said China's limited resources have spurred its researchers to be innovative, particularly through algorithm-hardware co-design, which enables AI firms to run large models on smaller, inexpensive hardware.

Tang Jie, founder of Zhipu AI which raised HK$4.35 billion in its IPO, also highlighted the willingness of younger Chinese AI entrepreneurs to embrace high-risk ventures - a trait traditionally associated with Silicon Valley - as a positive development.

"I think if we can improve this environment, ‌allowing more time for these risk-taking, intelligent individuals to engage in innovative endeavors ... this is something our government and the country can help improve," said Tang.


Brew, Smell, and Serve: AI Steals the Show at CES 2026

German group Bosch presented its fully automated 800 Series coffee machine (sold from $1,700) that can be synchronized with Amazon's Alexa voice assistant. Thomas URBAIN / AFP
German group Bosch presented its fully automated 800 Series coffee machine (sold from $1,700) that can be synchronized with Amazon's Alexa voice assistant. Thomas URBAIN / AFP
TT

Brew, Smell, and Serve: AI Steals the Show at CES 2026

German group Bosch presented its fully automated 800 Series coffee machine (sold from $1,700) that can be synchronized with Amazon's Alexa voice assistant. Thomas URBAIN / AFP
German group Bosch presented its fully automated 800 Series coffee machine (sold from $1,700) that can be synchronized with Amazon's Alexa voice assistant. Thomas URBAIN / AFP

AI took over CES 2026, powering coffee machines to brew the perfect espresso, a device to create your perfect scent, and ball-hitting tennis robots that make you forget it's human against machine.

Alexa, make me an espresso

German group Bosch presented a new feature for its fully automated 800 Series coffee machine (sold from $1,700) that can be synchronized with Amazon's Alexa voice assistant, said AFP.

After a short night's sleep, users can order a double espresso with voice commands only, and the coffee maker will deliver. Some 35 different espresso options are available.

"We're one of the first manufacturers to really lean in with AI," explained Andrew de Lara, spokesperson for Bosch.

The century-old company, positioned at the high end of the market in the United States, wants to gradually bring AI into the kitchen, notably through its Home Connect mobile app, which already allows users to control several appliances remotely.

Scent of AI

South Korean company DigitalScent has developed a machine, already available in some airports, that creates a personalized fragrance based on your mood and preferences.

Once you have picked your preferences, it releases a scent that gives you an idea of the final result. You can then make adjustments before making your final decision.

Once you have placed your order, the machine uses AI to produce a virtually unique fragrance in a matter of seconds, choosing from a range of over 1,150 combinations.

The fragrance is contained in a small, portable vial, costing $3 to $4, according to a spokesperson.

Game, set, AI

Several start-ups unveiled new-generation ball machines powered by artificial intelligence.

While Singapore-based Sharpa already offers a convincing humanoid table tennis robot with a reaction time of just two hundredths of a second, there is no equivalent on the market for tennis.

A few days ago, China's UBTech posted a video online of its Walker S2 robot playing rallies with a human, but at a slow speed and without any real movement.

UBTech's robots are designed for industrial use rather than tennis courts and, in all likelihood, the video was produced solely to demonstrate the agility of the Walker S2 to attract business customers.

While we wait for the humanoid robot that can volley at the net, another Chinese company, Tenniix, is marketing a robot that sends balls at speeds of up to 75 miles per hour (120.7 kilometers per hour).

It has 10 different shots, some with spin, and even a lob that reaches eight meters high.

The basic version, which can hold up to 100 balls, will set you back $699, but the most complete version, at $1,600, includes cameras and wheels that allow it to move around.

The fast-moving machine uses AI to analyze the trajectory of your cross-court forehand and fires off a ball from about where a real-life return shot would most likely come, giving the player the impression of a real rally.

"There's a real rhythm," says Run Kai Huang, spokesperson for Tenniix, "as if you were playing with a real person."


Award-Winning Game Studio Chief Rules Out AI Art

AI (Artificial Intelligence) letters and robot hand are placed on computer motherboard in this illustration created on June 23, 2023. (Reuters)
AI (Artificial Intelligence) letters and robot hand are placed on computer motherboard in this illustration created on June 23, 2023. (Reuters)
TT

Award-Winning Game Studio Chief Rules Out AI Art

AI (Artificial Intelligence) letters and robot hand are placed on computer motherboard in this illustration created on June 23, 2023. (Reuters)
AI (Artificial Intelligence) letters and robot hand are placed on computer motherboard in this illustration created on June 23, 2023. (Reuters)

The head of Larian Studios, the developers behind 2023's game of the year "Baldur's Gate 3", has vowed to ban any use of AI art in the outfit's upcoming project "Divinity".

The intervention by Swen Vincke follows repeated episodes of fan outrage over AI art in games in recent months -- with this year's game of the year winner "Clair Obscur: Expedition 33" stripped of its Indie Game Awards wins over alleged use of generative AI.

"There is not going to be any GenAI (generative AI) art in 'Divinity'," Vincke said Friday in an "Ask Me Anything" session on discussion site Reddit.

Fans had blasted Larian last month after Vincke told Bloomberg some generative AI was being used during development.

"We already said this doesn't mean the actual concept art is generated by AI, but we understand it created confusion," Vincke posted on Friday.

"To ensure there is no room for doubt, we've decided to refrain from using GenAI tools during concept art development," he added.

Vincke had said in December that the team's use of generative AI was "to explore references, just like we use Google and art books... at the very early ideation stages".

The new "Divinity" -- revealed to great fanfare at the December 11 Game Awards in Los Angeles -- is hotly awaited by gamers enthralled by the sprawling story and engaging characters of "Baldur's Gate 3", which has sold more than 20 million copies.

Despite his commitment on AI art, Vincke said that generative AI "can help" with other aspects of development, as studios "continuously try to improve the speed with which we can try things out".

He insisted that would benefit gamers through "a more focused development cycle, less waste, and ultimately, a higher-quality game".

Some executives believe generative AI's infusion into the industry will lead to a flowering of more ambitious titles that cost less to produce.

But they are running up against artists' fears that they will be pushed out of work and some gamers' concern that AI use will make for blander, less creative work.

French title "Expedition 33" saw its Indie Game Awards titles including "game of the year" withdrawn last year over some AI-generated art assets, which developers Sandfall Interactive insists were placeholders that it replaced in an update to the final game.