Thought #26 - Dependent, Distracted, Disrupted

Hi lovely humans,

The general sentiment of the AI news this week has felt a bit tragic (for the humans it is impacting). We’ve seen in full force just how dependent people are becoming on these tools with the reaction to GPT-4o being removed from ChatGPT. This week’s news included - AI induced psychosis, an AI convincing an elderly man to travel to New York to meet them and several other upsetting stories.

While we usually focus on using these tools for work, we are after all Taught by Humans, so the human side of any technology is important to us. This reminds me of the impact social media platforms have, particularly on teenagers (if you haven’t seen the Social Dilemma on Netflix, it worth a watch).

Hopefully this time, we’ve learnt from those mistakes and we’ll make efforts to protect people from the negative impacts of technology. Whether this is a call for regulation, greater accountability to these tech companies or education to keep people safe, I’m not sure (probably all three?).

In other news - lots going in the UK government and public sector with AI. Doctors reporting great time savings. People in the US are more worried about their jobs than before. A HR tech firm has seen a big data breach.

Our Week in AI

GPT-5

I’ve still not warmed to it. I am using it more, but I find I am still switching back to GPT-4o for any long form writing (as GPT-5 just isn’t giving me what I want - it sounds off still). The key issues I’m finding:

1. Doesn’t deal as well with my terrible typing and spelling mistakes

2. Taking things very literally, and unable to get nuance from conversation

3. Gets stuck in rabbit holes with ways of thinking

Examples - writing this newsletter (but it was excellent at proof reading), trying to find a nice acronym for our pillars of AI, helping me with a grant application.

Weirdly Playing AIs Off Each Other

In writing the grant, I was switching between Perplexity for research and ChatGPT for writing. I decided to get them both to write responses, and then get the other to critique. An unnecessary rabbit hole, but entertaining as Perplexity was super polite about ChatGPT, whereas ChatGPT did not return the sentiment.

What Actually Worked

1. I wrote bullet point notes for each response.

2. ChatGPT drafted something.

3. I used ChatGPT in temporary chat (so it doesn’t remember our other conversation) to mark the response after the mark scheme.

4. I also asked Perplexity to mark it.

5. Made the needed changes.

6. Used Perplexity to find the best references to support my claims.

I don’t find ChatGPT great at providing useful sources (the deep research mode is a little better), and I don’t find Perplexity good at chatting or refining. So using both is how I get things the way I want.

AI New Releases

Delightfully a bit quiet on the new releases this week (expected after everyone dropping so many last week):

• Claude launches two new ways to learn – Claude Code now has an /output-style command so you can switch between “Explanatory” (step-by-step reasoning, trade-offs, best practice) and “Learning” (a pair-programmer approach where you’re prompted to try tasks yourself). The “Learning” style is also being rolled out more widely in the Claude app - not just for education, but for anyone who wants support in understanding a tricky concept rather than being given the answer straight away.

• OpenAI adds connectors to Microsoft tools – ChatGPT can now pull in context directly from Teams, Outlook email, and Outlook calendars (in beta for Team, Enterprise and Edu users). The connectors are off by default for Enterprise and Edu, but admins can switch them on in Settings → Connectors.

AI News

• Workday data breach – Company confirms attackers gained access to some customer information.

• Meta’s internal chatbot rules under scrutiny – Reuters investigation finds their moderation guidelines allow harmful responses to slip through.

• 71% of Americans worried about job loss from AI – A new poll suggests concern about long-term displacement is rising.

• Meta restructures AI teams for the fourth time in six months – Constant reshuffling reflects pressure to catch up in the AI race.

• AI investment wave fuelling “zombie” companies – Interesting analysis on how inflated valuations and shallow deals could leave hollow firms in their wake.

• AI misidentification in suspected crime – Another cautionary example of risks in policing.

• Recreating faces of Welsh convicts sent to Australia – A curious use of AI in historical research.

Government and Public Sector

• AI to reduce doctors’ paperwork – Part of the UK’s ongoing drive to free up clinical time.

• Push for a sovereign UK AI chip industry – The government’s science advisers say Britain shouldn’t be reliant on overseas suppliers.

• AI career coaching – New pilots exploring AI guidance for jobs and even relocation.

• Liverpool schools improve exam results with AI – A BBC piece on how AI is linked to better outcomes in local schools.

• Wolverhampton hospital trust reports positive results from AI – Improvements across clinical services from an AI rollout.

• Jersey GP surgery seeing benefits from AI transcription – Small-scale but encouraging results.

• Study shows doctors can quickly become reliant on AI – Raises questions about long-term skills and judgement.

AI Companions

• AI companions and relationships – A rare positive story, alongside ongoing ethical concerns.

• Meta chatbot linked to man’s death in New York – Raises difficult questions about safety, responsibility and human decision-making.

• Reddit thread on “ChatGPT-induced psychosis” – Anecdotal but unsettling, part of a wider trend of mental health questions tied to chatbot use.

Not Quite News, But Worth a Read (or Listen or Watch)

• Guardian - Black Box Podcast - This isn’t a recent podcast, but I have found myself recommending it (especially episode 3 which is eerily similar to the GPT-4o removal reaction last week). It’s a good grounding in the history of AI, and very easy listening.

Help please - We’re oddly light on this section - if you’ve been reading any think pieces, or listening to a good podcast, please do share.

LinkedIn AI Poll

What do people think of GPT-5?

Apparently 1/3 of people don’t think it’s changed much (this was my original vote, now I’m not so sure). Nearly another third think it’s better. A very different response than what I was seeing on Reddit, so I can only assume people using it for work are much less impacted by personality changes than people using it as a companion.

Vote in this week’s poll - please!

This week we want to know if you are using open source AI models at all.

Blogs and Resources

OpenAI’s awful graph to demonstrate how powerful GPT-5 was has been talked about just as much as this AI breakthrough. Which got me thinking about whether data skills actually still matter.

Final Thoughts

As always we hope this was helpful!

Feel free to share this with anyone who might find it useful.

Next week, we are working on an open source series which I can’t wait to share.

Laura
Always learning