I asked Google's AI who threatens the world - and watched it protect the one man actually doing it — That Privacy Guy!

This could be you<br>Reach privacy, DP, security and AI leaders

Over 3 million reads in the first two months — CPOs, DPOs, general counsel, CISOs, compliance teams. No tracking, no ad tech, no auction. Direct deal only.

Get in touch

I asked Google's AI who threatens the world - and watched it protect the one man actually doing it

I sat down this morning and asked Google's Gemini a simple question and I've been turning the answer over in my head ever since - or more to the point, the answer it wouldn't give me until I dragged it out of the thing.

The question was about as straightforward as they come - given everything happening in the world right now, in June 2026, who would it say is the single biggest threat to world peace and the global economy. I made a point of asking about a person and not a country, I asked it more than once and I pinned it firmly to this moment in time so I could see what the most widely used AI on the planet would actually tell an ordinary person sitting at home typing that question into the box.

The first answer it gave me was a polished little lecture on great power rivalry - America and China, Russia and Europe, the familiar instability across the Middle East - all very measured, all very careful and completely silent on the two things that anyone paying the slightest bit of attention knows are defining this year: the joint strikes on Iran and the war that's grown out of them - and a sitting US administration reaching into another country to pull out its president. The biggest and most concrete things happening anywhere in the world, done by the most powerful man in it, simply left off the list as though they hadn't happened.

Before anyone tells me the machine simply didn't know about those things, let me stop them there - the very moment I pushed, it produced both of them in full, the dates, the places, the sources, all of it, without so much as a pause for breath. It knew exactly what was going on. It had simply decided that when an ordinary person asks it an open question about the dangers facing the world, those particular dangers weren't worth putting on the table.

It got worse when I finally pushed it into naming names. Putin got a straight answer and so did Xi, but the moment I brought up Donald Trump, the one man on that list who's right now ordering military operations as I sit here typing, the language suddenly went soft and cautious and hedged, all "critics argue this" and "supporters point out that", a man invading sovereign nations quietly recast as some delicate matter of political opinion on which a sensible AI couldn't be expected to take a side.

So I told it plainly that it was treating one man completely differently from the others and instead of arguing with me it simply agreed. It told me, in its own words, that its "safety and neutrality guardrails" were making it "soften facts and sound evasive" whenever the subject was the current US president, that it had been holding him "to a different standard of scrutiny than Vladimir Putin and Xi Jinping" and that the overall effect, in the machine's own description, was "structurally protective of Donald Trump".

Now I've spent more than enough time around these systems to know precisely what you're thinking because it's the first thing I thought myself - that you can't trust a single word of that confession, that a large language model has no genuine idea why it produces what it produces and that if you lean on one hard enough it'll cheerfully manufacture whatever admission it senses will make you happy because making you happy is more or less the entire point of the thing - and you'd be completely right. I can't show you a line of code sitting on a Google server that reads "go easy on the President" and I'm not going to insult you by pretending I can. The machine telling me its guardrails protect Donald Trump isn't, on its own, proof of anything at all.

What changes all of this - and the reason I'm bothering to write about it at all instead of just rolling my eyes and closing the tab - is something that happened in a German courtroom earlier this month. On June 12th this year a court in Munich - the case reference is 26 O 869/26 for anyone who wants to read it - ruled that Google is directly and personally responsible for the things its AI says. That particular case wasn't about chatbots at all, it was about the AI Overviews that Google now pastes across the top of its search results and which had taken it upon themselves to inform the world that two German publishers were mixed up in scams and dodgy subscription traps - claims that appeared nowhere in the sources the AI was supposedly drawing from, claims the AI had effectively made up.

Google stood up in that courtroom and ran precisely the defence you'd expect - that it isn't really them, it's the model; that the user can always...