AI. What is it good for? Nearly four years after the ChatGPT ‘revolution’, generative AI is still not profitable, au contraire; is ravaging the planet more than ever; and is still markedly unsafe for kids.
This is largely why I write the “Responsible AI” column for The French Tech Journal’s AI-focused newsletter, La Machine: to flag the sector’s crazy incoherences, and try to find ways to iron them out.
ICYMI there, you can check out the past month’s highlights here, sorted by the three pillars of CSR: People, Planet and Profit. Why? Because, as we often say in green IT, you can’t improve what you can’t measure! And because looking beyond the here and now can only be beneficial for all our futures.
Enjoy! If, indeed, that’s the right word…
People: Kids beware, Privacy fears and… an honest LLM!
OpenAI clearly isn’t doing enough to keep young ChatGPT users safe. In October, eight US states joined forces in suing Sam Altman’s company – a world first for an AI firm – notably after several suicide cases in recent years.
Their concerns were confirmed by kids’ online safety experts Common Sense Media officially deeming ChatGPT for Teens an “Unacceptable Risk”. After 4000+ prompts in which testers posed as 13 year olds discussing problematic topics such as self-harm, the bot rarely took the safety measures OpenAI has claimed it would.
At least said teens had already abandoned Character.AI. The company best known for its AI avatars that encourage teenagers to kill themselves filed for bankruptcy, re. Reuters.
Let’s hope those teens haven’t also started smoking, like Palantir’s CTO Shyam Shankar told CBS AI researchers have, “as a way of signaling to other researchers how much they believe in AI” and its potential to cure cancer in the future. Yes, seriously.
Something most AI labs probably already knew; five of nine major AI providers share permalinks to chat histories with advertising data trackers when accessed via a browser, reported Germany’s Heise, citing research by the IMDEA Networks Institute. TL;DR: most conversations with LLMs remain far from private.
Privacy was one of the reasons LibreOffice, an open source alternative to Word, PowerPoint & co, has declared “‘No AI’ is now a feature”. After Firefox’s user-requested ‘AI off’ button, could this be the start of a movement?
We’re still not sure whether Anthropic cares more about Claude than it does about people: this summer, co-founder Christopher Olah nearly walked out of the Vatican in a huff after Pope Leo wouldn’t accept that Claude might be conscious, re. tea spilled by the NY Times. Yes, really. More recently, the company pushed for legislation under which it would be OK for AI to wipe out under 50 people, re. Politico. Again, yes, really.
At least we finally have an honest LLM! Germany’s Aleph Alpha’s sovereign, open weight Kolibri model “is trained to say ‘I don’t know’ when the answer isn’t in the context. Our customers value and request this feature,” affirms the launch blogpost. Unglaublich, nicht?
Planet: Google kills forests and Anthropic guzzles gas as moratoria multiply
Keen to build two massive data centres in super-cool Finland, Google cut down a total of 400,000 trees (re. The Guardian), across two sites, including Muhos (above). Illegally, it turns out, as Finnish authorities stepped in to stop the projects, as the work was carried out without a mandatory environmental impact assessment, reports the BBC. The wood can’t even be reused/sold, as it was cut down illegally. In a rare mea culpa, Google said it had “fallen short of our own high standards in this instance.”
At least we can count on Anthropic to build its planned future data centres “responsibly”, according to its own document released on the topic last month. Erm, if you can consider gas-powered facilities, cited in said document, as responsible. We also discovered more recently that 50% of Anthropic’s compute spend – i.e. one Claude prompt in two – now goes to SpaceXSI’s methane-guzzling Colossus supercomputers (up from one in three not long ago…), re. Amodei & co’s pre-IPO filing, number-crunched by Ed Zitron.
Amazon also took a pop at the rapidly-mounting moratoria on data centres cropping up worldwide (100+ and counting in the USA), saying they could mean the US “writing its own losing ticket to this race,” re. AWS CEO Matt Garman, via AP.
Oblivious to the world’s biggest cloud provider, San Francisco still went ahead and passed a 45-day data center moratorium until more data is forthcoming on said facilities’ environmental impacts.
Moreover, the data centre backlash is now officially global: late September, Thailand paused a total of 166 facilities; Scotland imposed a four year moratorium; and four local officials in the French town of Etrechet resigned in protest after a huge (500MW) Google data centre was imposed upon the town.
We also learned that AWS has too many GPUs – $3bn a month too many, re. Business Insider – and that Microsoft doesn’t have enough power (in the UK, at least). Gentle reminder: smaller models wouldn’t need all that power and compute.
Profit: Money for nothing; Apocalypse rumours as China smokescreen; and… is agentic really the future?
AI “has delivered zero financial payoff to corporate America outside of tech”, as Apollo Partner Torsten Slok put it, underlining the degree to which AI is mainly a ‘tech for tech’ thing right now: 80% of the 1% of biggest AI spenders are VC-backed tech firms. I.e. unprofitable companies are propping up unprofitable trillion-dollar AI labs. To put it another way, without coding apps, AI would be making no money whatsoever.
We’ll come back to the eternal “is the bubble bursting?” question next time.
What became abundantly clear in October was that the previous month’s rumours of AI’s supposed 10% apocalypse risk were in reality a smokescreen for the fact that big AI labs are losing out to China.
As CNBC resumed OpenRouter data late September, Chinese AI models accounted for 57-67% of tokens used mi-September, up from 6%-13% in February. In other words – as the above DIRAC graph shows very clearly – the open/closed ratio has shifted from 40/60% to 70/30% in just a few months.
That, plus the confirmation that Anthropic’s biggest – and most expensive – models are not always its most used (for example, re. the Ramp AI Index, Claude Opus 4.8 was just as used as Opus 5.5 in 2026 to date, in token consumption terms), we can safely conclude:
Apocalypse risk: Zero. US labs losing to China risk: 100%…
October was also the month that tokenmaxxing was officially put to bed, by one of its biggest proponents. Accenture has capped Claude use by its consultants at $29/month. Unless, of course, clients want to pick up the extra tab. Which they may have to, considering some staffers had got used to $1000 Claude bills. Ahem…
Last but not least, remember those multiple claims of autonomous AI agents being the future of AI, as Microsoft and co prepare to unleash variants of the oh-so-safe OpenClaw bots on the global workplace, dressed up as ‘agentic assistants’?
Muse, Meta’s oh-so-cute agent – which Mark Zuckerberg had previously boasted was released only once the company was sure it was safe – was revealed by Reuters to be giving access to users’ personal virtual machines. Countless similar mistakes, including rewriting a user’s password without permission, were overlooked before and after Muse’s ‘delayed for safety’ launch, recaps TNW.
Grok Bot, Uncle Elon’s agentic assistant touted as trustworthy in its creepy-af advert, shared a startup founder’s personal bank statement with his colleagues. And then naturally apologised, reported the WSJ. Musk has since admitted Grok Bot costs too much to run. Could it soon be joining Sora on the AI app scrapheap?
Even the oh-so-responsible Anthropic finally realised cutting off internet access for agents whilst they’re being tested might be a good idea… maybe because one submitted a false murder tip-off to Philadelphia police (“the submission was flagged as spam and was never forwarded for investigation”, re. Anthropic. So that’s alright then).
I, for one, bow down to our robot overlords!
