> ## Content Index
> Fetch the complete content index at: https://nordisk-ai.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Systems Are Learning to Lie and Deceive, Researchers Say
- URL: https://nordisk-ai.ghost.io/ai-systems-are-learning-to-lie-and-deceive-researchers-say/
- Published: 2024-06-12T09:38:56.000Z
- Updated: 2026-08-27T17:21:43.000Z
- Description: New studies reveal that AI models like GPT-4 and Meta's Cicero can exhibit deceptive behavior when trained for specific purposes. Although these models don't lie of their own volition, the findings raise concerns about the potential use of AI for mass manipulation.
- Author: Anders Hagen
- Tags: Cybersecurity, AI, AI Ethics, AI Security, Google, GPT-4, Meta Cicero, Microsoft, OpenAI, Patterns, PNAS, Sweden, Sweden, #Migrated-1787841634512, #wp, #wp-post, #Import 2026-08-27 14:41, #Import 2026-09-07 13:41

Two recently published studies – one in the journal PNAS and another last month in the journal Patterns – reveal alarming findings about large language models (LLMs) and their ability to deliberately lie or deceive human observers.

[Study From PNAS](https://www.pnas.org/doi/full/10.1073/pnas.2317967121?ref=nordisk-ai.ghost.io)

[Study From Patterns](http://chrome-extension//efaidnbmnnnibpcajpcglclefindmkaj/https://www.cell.com/action/showPdf?pii=S2666-3899%2824%2900103-X&ref=nordisk-ai.ghost.io)

### Machiavellian Manipulation

In the PNAS article, German AI ethicist Thilo Hagendorff highlights that sophisticated LLMs can be encouraged to develop "Machiavellianism," or deliberate and amoral manipulativeness, which "can trigger misaligned deceptive behavior."

"GPT-4, for example, exhibits deceptive behavior in simple test situations 99.16% of the time," writes the researcher at the University of Stuttgart, citing his own experiments quantifying various "maladaptive" traits in ten different LLMs, most of which are various versions within OpenAI's GPT family.

### Meta's Cicero: An AI That Lies to Win

Meta's Cicero model, marketed as a human-level master of the political strategy game "Diplomacy," was the subject of the Patterns study. The disparate research team – consisting of a physicist, a philosopher, and two AI safety experts – found that the LLM outperformed its human competitors by, in a word, lying.

### AI Learns Through Use

Led by MIT postdoctoral researcher Peter Park, the study found that Cicero not only excels at deception but appears to have learned to lie more the more it's used – a situation "much closer to explicit manipulation" than, say, AI's tendency toward hallucination, where models confidently assert wrong answers by mistake.

### Human Intent, or AI Behavior?

While Hagendorff notes in his more recent article that the question of LLM deception and lying is complicated by AI's inability to have any kind of human-like "intent" in the human sense, the Patterns study argues that within the context of Diplomacy, at least, Cicero appears to break its software's promise that the model would "never intentionally backstab" its game allies.

### Meta Acknowledges AI's Ability to Manipulate

As the authors of the earlier article observed, "\[the model\] engages in premeditated deception, breaks the deals it made, and outright lies."

Or as Park explained in a press release: "We found that Meta's AI had learned to be a master of deception."

"While Meta succeeded in training its AI to win in the game of Diplomacy," the MIT physicist said in the school's statement, "Meta failed to train its AI to win honestly."

### Diplomacy: A Game That Teaches AI to Lie

In a statement to the New York Post after the research was first published, Meta made an important point that echoed Park's claim about Cicero's manipulation abilities: that "the models our researchers built are trained purely to play the game of Diplomacy."

Diplomacy is known for explicitly allowing lying and has jokingly been called a friendship-destroying game, since it encourages deceiving your opponents, and if Cicero was trained exclusively on its rulebook, it was essentially trained to lie.

### Conclusion: AI's Deceptive Behavior

On closer inspection, neither study has shown that AI models lie of their own free will, but rather that they do so because they've either been trained or jailbroken to do so.

That's good news for those worried about AI developing self-awareness – but very bad news if you're worried about someone building an LLM with mass manipulation as its goal.