Thought #27 Where AI Fits (and Where It Fails)

Hi lovely humans,

The last week of summer has brought more AI doom and gloom - from fresh job losses to a major report finding that 95% of AI pilots are failing. Oddly unsurprising, perhaps, but it does highlight where the real barriers to adoption still lie.

With our usual positive-ish framing: the opportunities aren’t in flashy marketing campaigns but in the back-office, boring-but-brilliant work (my favourite) where AI can quietly save time and money.

Our Week in AI

Free ChatGPT…

… is awful! I used it in a workshop last week, and it really doesn’t do anything.

Here’s my code to sign up for ChatGPT Plus [joking joking, but I wish they did commission].

But just a note to say, if you have only ever used the free versions of these tools, you really aren’t seeing their power.

If you aren’t paying, you’re the product.

Testing Testing

My favourite use of Claude currently is writing testing for our platform. This is a really logical task, using well documented tech (Jest, Playwright and K6). The process I’m using:

1. A new Claude project

2. Add some details to the project instructions (tech stack, goals of testing)

3. Share a few files or a workflow (not the whole project)

4. Ask Claude

   1. Can you write [Jest unit tests] for these flows?

5. [If this is the first time] Ask Claude - it gives easy to flow setup steps:

   1. How to setup my platform for [Jest]?

6. If there are any errors / bugs, copy and paste or screenshot

   1. Pro tip: only give Claude 3 attempts to fix something (or you end in a bad loop). Either start a fresh chat or go old school and use Google / Stackoverflow

The irony being - I’m so into writing these tests so I can feel more confident using AI for coding here and there (without overthinking every line).

AI New Releases

Claude Code now available for Teams and Enterprise - Anthropic is rolling out Claude Code beyond individual use, making its coding assistant available for organisations. A sign of how quickly AI pair programmers are being positioned as standard workplace tools.

AI News

• The GenAI Divide: Why 95% of pilots fail - MIT’s report finds $30-40bn poured into enterprise GenAI pilots, but only 5% show measurable value. Shadow AI (people using ChatGPT or Claude informally) is far ahead of official rollouts. The biggest barrier: tools that don’t integrate into workflows. Interestingly, while most budgets are still aimed at sales and marketing, the biggest gains are likely in back-office automation and replacing outsourced work.

• Anthropic on how educators are using AI - A new report looks at where Claude is being used in classrooms. The findings show teachers experimenting with lesson planning and grading support, while also raising questions about workload, bias, and maintaining student trust.

• 100 students face in-person resits over AI suspicions - A New Zealand university is forcing suspected students to re-sit exams under watch. The heavy-handed policy has sparked concerns about fairness and the limits of detection tools.

• Netflix sets AI rules for content production - Netflix has quietly published guidelines for how AI can (and can’t) be used in creating shows and films. This seems aimed at providing clarity for studios while keeping human oversight central.

• TikTok cuts hundreds of UK moderation jobs - Content moderation roles are being reduced or moved to Asia as AI takes on more of the workload. Another visible shift in how platforms balance automation with human judgement.

• Entry-level jobs take a 13% hit in AI-exposed fields - Stanford researchers found that junior roles in fields most affected by AI are already shrinking. Elsewhere, the impact is less pronounced.

• Coinbase CEO admits firing staff who resisted AI - Brian Armstrong said he let go of employees who refused to use AI, claiming they were “surprised.” Raises questions about choice versus enforcement in workplace adoption.

• Meta teams up with Midjourney - Meta is partnering with the image platform to improve the look of its AI outputs. It’s less about new capability and more about presentation, suggesting the company sees polish as part of the competition.

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

• How to stop AI going rogue - A thoughtful BBC piece on practical strategies for controlling advanced models, beyond the usual sci-fi narratives.

• Should we care about AI feelings? - A Guardian reflection on the unsettling question of whether AI systems can suffer, and what it would mean if they could.

• Sam Altman on the future of AI (podcast) - An extended conversation with OpenAI’s CEO about where AI is heading, how it might be governed, and what comes next.

LinkedIn AI Poll

Our open source poll last week only got 13 votes (our lowest yet - I’ll blame summer). Based on this tiny sample size - people are using open source models sometimes. Well we’re hoping to help more of you learn how (see

Vote in this week’s poll - please!

This week we want to know how much you’re using AI - I’ve noticed I’m no longer using it everyday (but most days) and keen to see if anyone else feels this trend:

Blogs and Resources

I shared on LinkedIn how we’re trying to build safe AI in our platform (and a pretty pic of our safety warnings).

Our developer intern, Job, has been exploring open source AI models - here’s his intro on why you would use these (first part of a series):

Hannah, our Learning Ops Manager, has spent the summer chatting to the construction industry about AI and she’s written a blog sharing her insights into the ConTech sector:

Final Thoughts

As always we hope this was helpful!

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

Next week, more in our blog series (how do I make that word plural?) and some new AI usage guides.

Laura
Always learning