🤖 AI News Roundup: Sloppenheimer, Fable 5, AI Sovereignty & Gary Marcus

The headline numbers say AI has won the workplace. Stanford's 2026 AI Index puts organisational adoption at 88%, and surveys of finance teams run higher still. The picture on the ground is harder to square with that, because engineers at Amazon and Google are mocking the tools they have been told to use, only a small fraction of organisations can point to real business impact, and the labs keep shipping ever larger models into the gap. It is the kind of mismatch that fuels talk of an AI market correction, and it runs underneath almost every story this week, from the lab releases through a regulatory shift to the slow unwinding of the outsourcing model.

The adoption-impact gap: workers reject mandated AI as "slop"

Engineers at Amazon and Google are openly mocking the AI tools their companies have mandated. Internal Slack channels overflow with jokes about failed adoption drives, staff dismissing AI output as "slop" and coining terms like "Sloppenheimer" to ridicule the tools they are expected to use. This is not a fringe view, it is the default reaction when organisations push adoption from the top down without proof of value.

Yet adoption headlines suggest momentum. Stanford's 2026 AI Index reports that 88% of organisations have implemented AI, and Gartner finds that 84% of finance organisations have implemented or plan to implement it. The striking number is the one underneath: only 7% of those finance organisations report high or very high business impact. The gap between adoption and measurable value has become the story organisations are not talking about, because it exposes something uncomfortable about how they roll out AI in the first place.

The issue is not AI itself, it is rollout discipline. When organisations mandate a tool before validating that it measurably improves the work, adoption becomes an exercise in compliance rather than value, and engineers and knowledge workers know the difference. They find flaws in tools they did not ask for and they say so in writing. The move for leaders is not faster mandates or better change-management comms, it is evaluation first: run real pilots with teams who care about the outcome, measure what genuinely improves, then scale to the teams where the value is clear. Let adoption be pulled by those who find genuine value rather than pushed from above.

Anthropic releases Fable 5, a Mythos-class model made safe for general use

Anthropic released Claude Fable 5, a Mythos-class model it describes as made safe for general use. The more capable Claude Mythos 5 itself stays restricted to Glasswing partners and, soon, select biology researchers, so this is the Mythos capability bracket reaching the public in a safety-constrained form rather than the unrestricted model being opened up. The release came with a new set of safety guardrails covering areas such as cybersecurity and the life sciences, alongside data-handling terms for business customers.

The guardrails are where the friction has landed. Cybersecurity researchers report that routine, legitimate work, including code review and writing secure code, is being caught and blocked by the safety classifiers, which struggle to tell defensive security work apart from misuse. The data-handling terms attached to the business tier are a further point of hesitation, with organisations weighing the retention conditions before rolling the model out widely.

For a worker or a small business, the practical question is not whether the capability is impressive but whether it fits the work in front of you. If your tasks touch security research or anything the classifiers read as sensitive, expect to hit blocks or fallbacks and budget time to work around them. For general writing, analysis, and software work, Fable 5 is a clear step up from what was previously public. Test it against your own use case, find where the friction sits, then decide whether the capability justifies the constraints.

Google scales text generation past 1,000 tokens per second

Google has released DiffusionGemma, an open-weights text model that generates roughly four times faster than the standard approach. It uses diffusion-based parallel decoding to reach more than 1,000 tokens per second on an Nvidia H100. Rather than producing one token at a time in sequence, the model drafts a block of text in a single pass and then refines it over several passes, working much as image diffusion tools do by starting rough and sharpening towards a final result.

Speed is not the point in itself. Inference cost scales with latency and throughput, so a model that generates four times faster can do the same job on a fraction of the hardware. A task that took 100 GPU-seconds might take 25, or a workload might run across fewer machines. For any business already weighing the per-token economics of putting AI into products and internal tools, that is the lever that matters, and it bites hardest at smaller scale where per-token cost drives the margin on every feature.

Europe writes its own definition of cloud and AI sovereignty

On 3 June the European Commission presented its Tech Sovereignty Package, and the piece that matters most for buyers is the Cloud and AI Development Act. Rather than a single blunt rule, the Act introduces an EU-wide framework for assessing how sovereign a cloud or AI service is in practice, streamlines the deployment of datacentres across the Union, and sets a target of at least tripling Europe's datacentre capacity within five to seven years.

The substance is a graded sovereignty test that a buyer can hold a provider against. The lowest level requires only that data is processed and stored inside the Union; the next adds independence from third countries and transparency over the software supply chain; a higher level requires the provider itself to be owned and controlled from the EU, with criteria reaching as far as the citizenship of its personnel; the top level demands full control of the supply chain and no interference from any third country. Public bodies pick the level their own risk assessment calls for rather than applying one standard to everything.

For an organisation in Europe or the UK, this turns a vague worry about where data lives into a graded question a procurement team can put to a vendor in concrete terms. A business running sensitive or regulated workloads can name the level it needs and ask its cloud or AI supplier to meet it, and a provider owned outside the Union cannot reach the upper rungs by definition. Sovereignty stops being a slide in a vendor pitch and becomes a line item in the buying decision.

Automation is unwinding the economics of outsourcing

For three decades, outsourcing rested on one principle: take work that can be defined, standardised, and monitored, then move it offshore and run it at lower cost. Generative AI is dismantling that model by automating the routine, rules-based tasks that justified the offshore business case in the first place. The work that made labour arbitrage profitable is increasingly the work that gets automated.

This is a structural shift rather than a cyclical one. When competitive advantage moves from wage differentials to automation capability, the geography of work stops being about who is cheapest and starts being about who has access to AI infrastructure and the technical talent to use it. The same logic that once pulled work towards low-cost regions now pulls it towards wherever that capability lives.

Roles in business process outsourcing, data entry, basic coding, and back-office operations are the most directly exposed, but the deeper change is in workforce composition. Organisations built on offshore labour arbitrage face a choice: retrain workers into the roles automation cannot reach, such as stakeholder management and complex judgment, invest in their own AI tooling to compete, or accept that the strategic value of the role has changed. That last option reaches workers in every geography, not only offshore.

When AI oversight becomes theatre

As organisations push autonomous agents into their workflows, a gap is opening between policy and practice. MIT Sloan's panel of practitioners describes a familiar failure: approval steps that were meant to be genuine checkpoints have turned into speed bumps. People asked to sign off quickly have no real chance to engage, so the box gets ticked, the liability is nominally transferred, and the work moves on. Oversight becomes performance rather than control.

That is arguably worse than no oversight at all. Agents can produce flawed analysis, incomplete recommendations, or biased decisions at scale, and a human skimming a summary before approving creates a paper trail that looks like diligence while delivering none. The risk is not the agent's autonomy in itself but the organisation telling itself it is still in control while quietly handing that control over.

For leaders deploying agents, the useful move is counter-intuitive: make the approval slower and more searching than the agent is fast. An agent racing through work that nobody meaningfully reviews is not acceleration, it is a liability waiting to surface. Set real verification checkpoints where the stakes are highest, such as financial decisions, customer commitments, and hiring, and train reviewers to interrogate the output rather than wave it through. If the approval is always fast, the oversight is not real.