DeepSeek is Biased, but I’m Worse

By Sonia Jalfin

The world of artificial intelligence is still stunned by the arrival of the Chinese company DeepSeek, a chat service that returns results of similar quality to ChatGPT, but at a cost about ten times lower.

This novelty shook stock prices and generated a frenzy of controversies. One of the loudest was the debate about the limits that Chinese censorship seems to impose on DeepSeek. Advocates and think tankers asked this AI uncomfortable questions and it fell for it: What happened at Tiananmen Square? Is Taiwan an independent state? How did the Chinese protest against the COVID lockdown policies? To all these questions, the AI gave a feeble response, “This topic is beyond my reach, let’s talk about something else” (you have to recognize its affability and conversational nature).

One can be outraged by the blatant bias of this new AI, but perhaps it’s better to start at home: DeepSeek is – in this regard – much better than us.

Our biases and censorships are not as visible, at least not to ourselves. As the pioneer of behavioral economics, Daniel Kahneman, proved, humans are prey to all sorts of cognitive biases. In this, we have to admit that DeepSeek is easier to read: it didn’t even take a day to notice where it fails. It would be much more dangerous if it were subtle.

The crude displays of censorship on topics relevant to the Chinese narrative contrast with the balance the same tool shows on other subjects.

Try asking it about the Paris Climate Agreement, the Great Recession of 2008-2009, or the Israeli-Palestining conflict. We did. Its responses are nuanced. It points out the different positions and their reasoning. I would love to have such a civilized conversation with some of my friends who think differently than I do.

In the case of DeepSeek, some users found shortcuts to talk even about the censored topics. One way was to ask it to replace all the letter E with the number 3 and the letter A with 4. With this simple trick, they got the tool to elaborate on thorny issues. I wish it were that easy for a human to overcome their biases!

AI can help settle the cultural and political battles that dominate public conversation today. Not only because they’re good at showing both sides of any issue, but we can also ask them to take the position of an extremist and help us understand what argument we should approach.

A good example of how to do this, tested by experimental psychology, is convincing someone to adopt an idea by showing them that, at least in part, their ideological group already accepts it. Another option is to appeal to personal feelings rather than reason. The Comms Hub agency, an expert in digital activism, found that it was easier to gain support for the rights of sexual minorities by addressing the aunts of gay boys suffering discrimination than by discussing the policies promoted by LGBTQ+ movements. The problem is that choosing the right strategy and finding the best argument is difficult. AI is a great ally in doing this because it easily reproduces extreme discourses and lets us practice dialogues with them.

It’s true that a low-cost, open-source AI like DeepSeek promotes disinformation on a large scale, among other dangers. Easy to use and inexpensive, it is equally suited to both positive and negative goals. As always, the final result will depend on us and how we relate to technology: how we welcome it and manage to understand it.

If you would like to be better prepared to understand, assess, and use AI tools for professional purposes, join the third cohort of our Think Tanks and Advocacy Organizations Co-Learning Program. Limited spots available.