25/ The effort to purge “woke AI” might force the hand of tech companies, leading to AI that is even more broken. This could saddle the military with faulty intel that harms soldiers and civilians. For the rest of us, we would be stuck with technology that leaves us more divided and less informed.

24/ OpenAI and Anthropic are selling their foundation models to the military.

Google is working with Lockheed to develop AI weapons.

Meta has bought a 49% stake in Scale AI, an up-and-coming defense contractor.

Amazon and Microsoft are major military cloud providers.

Billions are at stake.

23/ This time, tech company leaders have tried to pander to the administration’s “anti-woke” agenda, but this strategy could put them in a bind, particularly as they seek to land lucrative defense contracts that would offset the enormous costs of their AI business.

fortune.com/2025/01/13/z...

22/ Amazon complained this was retaliation against Jeff Bezos – Trump was unhappy at the time about reporting by WaPo, which Bezos owns. The suit was later dismissed after the contract was revised, and Amazon was awarded a portion of it.

www.reuters.com/legal/govern...

21/ This wouldn’t be the first time the Trump admin has tried to kill federal contracts in response to perceived slights. In 2019, Amazon sued the first Trump admin for using “improper pressure” to deprive it of a $10 billion military cloud computing contract.

www.nytimes.com/2019/12/09/t...

20/ But the wild card here is the administration itself – a chatbot response goes viral, Trump takes offense, and threatens to cancel the provider’s federal contracts – sounds familiar?

www.wsj.com/business/tru...

19/ The order provides room for this, clarifying that agencies should “account for technical limitations” in enforcing compliance and “avoid over-prescription and afford latitude for vendors.”

18/ Of course, tech companies could treat the order as symbolic, and placate the admin with messaging about anti-woke measures rather than meaningful changes to how their models are developed and used.

17/ The White House guidance does not apply to national security agencies, and the order allows for exceptions “as appropriate” for nat sec uses of Large Language Models. But the intent to mold foundation models in the admin’s image of what “truths” are acceptable is clear.

16/ This risk is not theoretical: in April, the White House updated its guidance on federal use and acquisition of AI to remove all references to bias detection and mitigation, including measures to account for the tech’s impact on underrepresented populations. www.lawfaremedia.org/article/narr...

15/ While the order is limited to Large Language Models, it’s easy to see how defense contractors could be pressured to strip out bias detection the admin doesn't like across foundation model training. This may undermine other military functions that increasingly rely on AI, like target recognition.

14/ If federal providers of AI-based translation services are prevented from fine tuning their models to be more accurate for languages other than in English, or to limit the reinforcement of harmful stereotypes, this could compromise the analysis of intelligence in these languages.

13/ The problem is that this could harm model performance. And the federal agency that this would impact the most is the government’s biggest spender on AI - the Department of Defense. www.brookings.edu/articles/the...

12/ Tech companies may also try to comply by rolling back internal measures to mitigate the spread of content that denigrates or erases racial and other minorities, or account for underrepresented languages in training data.

Critical studies like this could be shelved: openai.com/index/evalua...

11/ But fine-tuning is more insidious than content filters, since it biases the model to favor certain outcomes. We could end up with models that systematically produce biased responses in a neutral-sounding way, e.g. that there is no scientific consensus on the effects of climate change.

10/ This type of “fine-tuning” is apparently used to prevent chatbots from teaching users how to build bombs, conduct cyberattacks, or engage in dangerous or illegal action. Again, not foolproof - users can manipulate prompts to evade these restrictions.

www.theguardian.com/technology/a...