HAPPY FRIDAY,
Early July, which means Q3 is officially underway, the summer slide is setting in, and everyone is pretending they are not already half-checked out while simultaneously having their most anxious budget conversations of the year. Good time for a corporate espionage lawsuit, then. Apple is suing OpenAI for trade secret theft, and the allegations are specific enough to make you put down your coffee.
In This Week's AI Fridays:
- Apple vs. OpenAI: poached engineers, stolen hardware secrets, and a lawsuit with names attached
- Anthropic built a tool to see inside Claude's thinking, and some of what they found is a bit unsettling
- AI adoption in CX is at 90%, but nobody agrees on how to do it
- Research confirms AI fiction is detectable mostly because it is not very good
- Plus: Canva goes enterprise, model dependency risk, and a handful of quick hits worth your time
Apple Is Suing OpenAI for Allegedly Stealing Hardware Secrets
Apple's lawsuit alleges a specific pattern: former Apple engineers, reportedly encouraged by OpenAI, brought confidential presentations, secret prototypes, and key supplier relationships with them when they left. Senior OpenAI leadership is named as being aware of, or involved in, the alleged misconduct. That is a meaningful escalation from the usual inter-company friction over hiring.
The Highlights:
- Apple alleges OpenAI systematically recruited employees who brought confidential hardware materials across with them
- The suit implicates senior OpenAI leadership, not just individual engineers acting alone
- Stolen materials reportedly include secret prototypes and supplier details, suggesting OpenAI's hardware ambitions are further along than publicly acknowledged
- This lands at a moment when both companies are competing for the same device-layer real estate in the AI stack
The Takeaway:
OpenAI has spent years positioning itself as a research lab that became a product company. A lawsuit alleging it also ran an intelligence operation against its closest hardware partner is a different kind of story, and one that will be considerably harder to manage than the usual competitor noise.
Anthropic Found a Hidden Space Where Claude Puzzles Over Concepts
Anthropic built a tool called the Interpreter that offers the clearest view yet of what is happening inside a large language model as it works through a question. Some of the findings are roughly what you would expect. Others are not.
The Highlights:
- The Interpreter maps internal model activity as Claude processes prompts, giving researchers visibility into reasoning steps that were previously opaque
- Some findings are mundane; others fall into the category Anthropic researchers apparently describe as unnerving
- The tool represents a meaningful step forward in interpretability research, which has been something of a slow-moving field relative to the pace of model deployment
- Findings raise fresh questions about what "reasoning" in these systems actually constitutes, and how confident we should be in outputs we cannot fully explain
The Takeaway:
We have been deploying these systems at scale for years while having a fairly thin understanding of what they are doing internally. The Interpreter does not resolve that, but at least Anthropic is asking the question with more rigour than most. The unnerving findings are worth paying attention to.
AI in Marketing
AI Adoption Hits 90% as CX Deployment Paths Diverge
The adoption number is striking enough that it tends to end the conversation before the more interesting part starts. Yes, nine in ten organisations report using AI in customer experience in some form. What they cannot agree on is architecture, governance, or how to build customer trust while demonstrating any return on the investment. The divergence is sharp enough that two companies both claiming 90% adoption could be running almost entirely different programmes. Pete's read: the adoption metric is now largely useless as a competitive signal. What matters is what decisions the AI is making, how those decisions are governed, and whether customers know or care. Most organisations are still working that out.
AI Fiction Is Easy to Detect Because It's Stupid and Bad, Research Finds
The research did not find that AI fiction is detectable through forensic stylometry or sophisticated pattern analysis. It found that it is detectable because it is genuinely poor. ChatGPT has an apparent fondness for dream sequences. Gemini describes characters with something approaching obsession. Both tendencies show up reliably enough to function as fingerprints. This matters for anyone in content or brand work who is tempted to run AI-generated creative through a light edit and call it done. The tells are not subtle watermarks, they are craft failures, and audiences notice craft failures even when they cannot name them. The implication for agencies and brand teams is fairly direct: AI as a drafting tool is one thing, AI as the author of anything meant to resonate is still a different and considerably riskier proposition.
Canva Targets Enterprise Creativity with Trusted AI Creative Workflows
Canva has been the tool people explain apologetically to designers for years, and it is now making a deliberate push into enterprise territory, repositioning around editable, collaborative AI creative workflows with security and compliance built in. The pitch is less about generation and more about governance, which is the right instinct for enterprise buyers who have spent the last eighteen months watching their teams use consumer AI tools in ways that made legal nervous. Whether Canva can hold that positioning against Adobe and the purpose-built enterprise creative platforms is a genuinely open question, but the strategic read is sound. Enterprise teams do not need more AI that produces things. They need AI that produces things they can use, approve, and audit.
Protecting Your Work as AI Models Rapidly Come and Go
The model you built your workflow around six months ago may not be the model you are running on in six months' time. Models get deprecated, APIs change, pricing shifts, and the vendor you anchored your process to gets acquired or pivots. Marketing AI Institute raises the vendor dependency question with some seriousness here, and it is worth thinking through. The practical implication is not to avoid AI tools, it is to build workflows that are model-agnostic where possible, document your prompt logic somewhere portable, and resist the temptation to treat any one model as permanent infrastructure. We are still in the part of this story where the tools change faster than the organisations using them.
Quick Hits
- Apple's suit against OpenAI gets more specific the more you read it: poached engineers reportedly arrived carrying confidential presentations, secret prototypes, and supplier details. A habit, not an accident. Read more
- OpenAI launched ChatGPT Work, an agentic mode that autonomously acts across connected apps, files, and the web. Alongside the global rollout of GPT-5.6, which is apparently a thing now. Read more
- Benedict Evans on token pricing: AI is in a supply crunch today, but the more interesting question is what happens to pricing and margins when supply catches up. Spoiler: commodity infrastructure tends to get priced like commodity infrastructure. Read more
- As AI moves up the stack away from raw compute, it creates new and substantial enterprise lock-in risks that neither the critics nor the boosters are spending much time on. Worth reading before you sign anything multi-year. Read more
- Patreon is blocking AI training crawlers, with CEO Jack Conte citing creator credit, compensation, and consent as non-negotiable prerequisites. Notably, they are doing it rather than writing a blog post about doing it. Read more
- The AI infrastructure scramble has shifted: getting access to compute is now easier than using it well. Enterprise leaders are finding that utilisation, not acquisition, is the hard problem. Read more
This article was obviously generated with AI but curated by a human, don't be weird about it.
Compiled by
Pete Bishop
Chief Innovation Officer, ZGM Modern Marketing Partners
Pete Bishop has spent the last two decades helping brands adopt new technology without losing the plot. He hosts Artificial Breakdown, a podcast and weekly newsletter that translates AI news into practical marketing decisions, and scans 50+ sources each week to write this issue.
Get this in your inbox
One email a week with the AI and marketing news worth your time.