Politicians Trying to Edit the Robots Instead of Their Record

Saw a story this week about politicians reportedly trying to influence what chatbots say about them. Not through policy, not through actually doing a better job, but through some combination of astroturfing and complaint letters to AI companies, hoping the model’s answer about them shifts from “under investigation” to “champion of the people.” I read it twice because I wanted to make sure I wasn’t misunderstanding the ambition involved.

The instinct makes sense once you sit with it for a second. For twenty years the game was SEO. Get your press releases ranked above the bad news, bury the court documents on page four of search results where nobody goes. Now the bad news doesn’t even need burying if you can convince ChatGPT or Gemini to summarise your career in flattering terms when someone asks “who is this person.” One user in the comments put it better than I could: with search results you can usually tell what’s an ad. With a chatbot answer, you often can’t tell if it’s been shaped at all. There’s no sponsored tag on a hallucination.

What gets me is the sheer laziness of the approach. If your record’s bad, the actual fix is to stop doing the thing that made the record bad. That’s hard, though, and slow, and involves things like resigning or changing policy. Gaming a language model is comparatively cheap, and if you flood enough forums and comment threads with a preferred version of events, some of that sloshes into the next training run. Someone mentioned a subreddit that got a bot to claim a sitting politician had died. Nonsense, obviously, but nonsense is exactly the kind of noise that ends up as junk data floating in some future model’s weights. It doesn’t need to be believed. It just needs to exist somewhere the crawler goes.

I do IT for a living, mostly DevOps, and I spend a fair bit of time thinking about where data comes from and how much you can trust a pipeline you didn’t build yourself. Training an LLM on “the internet” was always a bit like building a house out of whatever timber washed up on the beach. Some of it’s solid. Some of it’s driftwood with a rusty nail in it. Politicians have apparently noticed that if they throw enough driftwood into the water upstream, some of it washes up in the pile.

There’s a genuinely hard problem buried under the cynical bit, and I don’t think it gets talked about enough. Who decides what a chatbot says about a public figure is a real governance question, not just a PR arms race. Too much intervention and you get a model that reads like a press release. Too little and the model becomes whatever gets gamed the hardest, by whoever has the most bot farms and the least shame. Neither outcome is “neutral,” whatever that word is supposed to mean when you’re talking about a system trained on humanity’s collective mess.

I don’t have a tidy answer here. I’m fascinated by what these models can do and genuinely unsettled by how much invisible shaping already happens before any of us type a question into the box. My daughter uses one of these tools for schoolwork most days, and I’ve caught myself wondering, not for the first time, whether the answer she’s getting reflects reality or reflects whoever spent the most effort making sure it did.

The comment that stuck with me was the one pointing out that ask any model a specific factual question and its “opinion” often shifts away from the polished version toward something closer to the truth, because the truth has more supporting evidence sitting around online than the spin does. That’s mildly reassuring. Evidence still tends to outweigh noise, at least for now, at least for things that are actually documented. Whether that holds as the astroturfing budgets grow and the noise gets cheaper to produce, I genuinely don’t know. I’d like to think reality has a home ground advantage. I’m just not willing to bet the house on it.