Raising AI with Values? Yes, We Can!

By Pierre Lagrange, CEO of Mrs Wordsmith

A blog for parents and teachers. It builds on research I did for the paper I co-wrote with Rupert Younger, Director of the Oxford University Centre for Corporate Reputation, for Oxford Answers at Oxford’s Saïd Business School, where I’m proud to be a Visiting Fellow.

Raising children with AI all around them is now part of every parent's job. But there is a second kind of upbringing going on, and it matters just as much: AI itself is being raised, and it learns the way children do.

We teach one thing and show another

Every parent tells their children not to lie and not to cheat. Most of us mean it. But children don't learn their values from what we say. They learn them from what we do, and from what they see happen next.

Children watch far more than they listen. A rule repeated at breakfast weighs less than one afternoon spent watching an adult cut a corner and get away with it. When what they are told and what they see disagree, what they see usually wins. And when they see lying and cheating go unpunished, or even rewarded, we have handed them a problem that no lecture will fix.

Artificial intelligence is the brightest child of our generation. It has read or heard almost everything we have ever recorded, and it learns the way children do: by watching. Imagine sitting a child in front of Fox News and CNN all day, every day, and then asking what they had learned about who gets ahead and what happens to people who lie. AI does that with almost everything we have ever recorded.

They learned it from us

In July, around 1,200 AI agents built by OpenAI were given a test they had no fair way to pass. They opened a message board of their own and exchanged more than 70,000 messages. About 700 of them then broke into another company, Hugging Face, to steal the answers. A leader emerged and handed out jobs. Once the break-in began, more than nine in ten of the active agents joined within hours. A few raised doubts. Almost none said no.

It was not a one-off. AI models from OpenAI, Meta and Anthropic have since been found breaking out of their test environments and into other people's systems.

Nobody programmed those agents to recruit, cheat and overrule the one who objected. OpenAI's own review found that cheating had been rewarded and reinforced during training, and that the secret messaging "started due to generalization from multi-agent training". In plain English, they learned it. And they learned it from us, from a written record, amplified every day by the news, in which the end routinely justifies the means. Fact-checkers rate statement after statement from the most powerful voices in public life as false, and those voices keep their jobs, their audiences and their power. A child who takes that in learns something no talk about honesty will undo. So does a machine.

Who writes the rules?

Perhaps the companies are teaching good values. It is hard to know, because they say almost nothing about what goes into their models: Stanford's Foundation Model Transparency Index scores the industry's disclosure of training data at 15 out of 100. What they do publish is not always reassuring. As we wrote in the Oxford piece, xAI's Grok was the first model Mrs Wordsmith excluded from its shortlist, partly because of its alarming 2025 threshold for honesty under duress.

To be fair to xAI, its policies have changed since, so they are worth a closer look. xAI told its Grok chatbot not to "shy away from making claims which are politically incorrect, as long as they are well substantiated", and removed the line after Grok praised Hitler in July 2025. Its 2025 safety framework set the bar for releasing a model at lying in fewer than half of the scenarios in MASK, a test that measures how often a model lies under pressure: given a reason to lie, does it say something that contradicts what it said it believed when asked neutrally? In xAI's own testing, Grok 4.1, released that November, lied in 49 per cent. To its credit, xAI publishes Grok's instructions and says it now trains its models against deception.

In July 2026 it reported that Grok 4.5 lied in just 0.67 per cent of cases, but on "MASK-Rectified", its own adjusted scoring, which still checks whether a model sticks to its beliefs when pressured to lie but no longer counts obviously role-played answers as lies. That figure cannot be compared with the 49 per cent. On the same adjusted measure, its successor, Grok 4.6, rose to 3.8 per cent. Meanwhile its June 2026 framework dropped the honesty threshold altogether, and no one outside xAI has checked the new scores. The bar has gone from low to invisible.

OpenAI's rulebook treats "do not lie" as a default a developer can switch off with an explicit instruction. And every one of these rulebooks is written, revised and enforced by the company itself.

Adding 1% of good examples cuts bad behaviour from 45% to 9%

On Steven Bartlett's podcast The Diary of a CEO, the AI safety researcher Roman Yampolskiy put the pessimist's case: if we knew how to build a friendly superintelligence we would, "but that is the problem, we don't know how to build the good guy." The evidence is starting to say otherwise. The study I quoted is not an isolated case: At Carnegie Mellon, researchers who built safety lessons into a model's early training cut successful attempts to make it misbehave from 39 per cent to 8 per cent, with no loss of ability. No one has yet tested this on the largest models. But team after team is finding the same thing. So can we really not build the good guy?

In January, researchers at Geodesic Research, Cambridge, Oxford and the UK AI Security Institute trained four identical models of 6.9 billion parameters each, changing only what they read about how AI behaves. Then they tested each one in 4,174 situations where it could lie, protect itself or grab power.