Governance Isn't Only a Brake. It's Also a Steering Wheel.

Good guardrails decide where AI goes in your business, not just whether it goes.

When most leadership teams hear "AI governance," they picture a brake pedal: approvals, committees, a list of don'ts. Some of that is necessary. But if braking is all your governance does, you're missing the more useful half.

A 2026 research report from Google DeepMind–affiliated authors makes the point in one sentence:

"It is worth noting that governance is not purely a brake on AI progress, but can also serve as a steering mechanism that shapes the direction and quality of development." (p. 27)

Note the wording: "not purely" and "can also." The authors aren't saying governance never slows things down. They're saying it does more than that.

What this report is, and isn't

The report is From AGI to ASI, an arXiv preprint by 14 named authors. All list a Google DeepMind affiliation, and a few also list universities. It maps possible pathways from human-level AI to superintelligence, and the "frictions" that could slow them. It is not Google policy (Google's corporate policy paper, from our earlier post on it, is a separate document), and it is not a forecast. The authors say so directly: "The future is unpredictable." (p. 2)

Governance is just one of several frictions the report discusses, but it's the one that matters most to operators.

Brake and steering wheel

The report names the kinds of tools it has in mind when it describes steering:

"Concrete instruments now in force or under development include compute-threshold-based licensing (European Union, 2024), mandatory pre-deployment evaluations and incident reporting (The White House, 2023), and international norm-setting through multilateral declarations (UK Government, 2023)." (p. 27)

That list reflects the authors' framing. The 2023 U.S. executive order it cites was rescinded in January 2025. The EU AI Act uses a training-compute threshold to identify the most powerful general-purpose models and puts obligations on their providers, such as model evaluations and serious-incident reporting, rather than running a licensing scheme as such. Rules here are moving fast.

It describes thresholds and evaluations as "embedding institutionalised gates that must be passed before additional capability scaling is politically and legally permissible" (p. 27).

Those are society-level instruments, and which of them apply to you is a question for counsel. But their shape is useful inside a company: a gate you pass before you scale up, a test you run before you deploy, and a record you keep when something goes wrong. None of those stops adoption. They decide which way it goes.

Why steering early beats stopping late

The more sobering point is about accidents. Summarizing research on complex-system failures, the authors note they "often originate in organisational decisions rather than isolated technical faults and can undermine social licence for further deployment" (p. 27). And:

"large, visible accidents or credible near-miss events could shift public preferences, liability regimes and regulatory thresholds in ways that render major further scaling steps towards ASI politically, legally or commercially infeasible, even where they remain technically and economically achievable" (p. 27)

The report means civilizational scale. My translation for a business (my framing, not the paper's): one visible AI failure with a customer can do more to stall your AI program than any policy ever would. A steering wheel helps you avoid the ditch.

The report also makes a related point about usefulness. Its authors write that "alignment difficulties may act at least to some degree as a direct bottleneck to capability development itself, as unsafe or uncontrollable systems cannot be well utilized for automated research or deployment." (p. 37) That's framed as an open research question, not a finding, and the report explicitly assumes safety gets solved so it can focus elsewhere (p. 37). The operator version is intuitive: a tool you can't trust is one you can't hand to your team.

The honest counterweights

The report doesn't treat this as one-sided. On slowing down, it notes that "others argue that capping progress rates may carry opportunity costs, like delaying advances in medicine, climate science, and economic productivity, which could also have a destabilizing effect on societies." (p. 27) Too much brake has real costs.

It also flags a limit: "asymmetric adoption of safety norms across competing nations could create regulatory arbitrage, where development migrates to jurisdictions with weaker oversight" (p. 28). There's an internal version of this: if your official AI process is too slow, people use unapproved tools, and your oversight covers less, not more. That's my analogy, not the report's claim.

Practical takeaways

These are good operating habits, not a legal checklist, and following them doesn't by itself mean you meet any regulation.

1. Write gates, not just bans. For each AI use case, define what has to be true before it scales: a successful pilot, a named owner, a reviewed data flow. "Not yet, until X" steers better than "no."
2. Test before you deploy. Run a short pre-launch check on anything customer-facing: sample outputs, edge cases, a human review of what could go wrong.
3. Keep a simple incident log. When an AI tool gets something wrong, write it down. Record what happened and what you changed, and keep customer personal details out of it. Near-misses count. That record is how your guardrails get smarter.
4. Watch for internal "arbitrage." If people are going around your process to use unapproved tools, treat it as a sign your process is too slow, not just a discipline problem.
5. Get advice on the external rules. Whether any AI-specific regulations, including state and industry rules, apply to your business depends on where you operate and what you do. Ask a qualified attorney.

This post is general commentary for business operators. It is not legal advice. Talk to a qualified attorney about your specific situation. The author is not a lawyer, and reading this doesn't create an attorney-client relationship.

Sources

• Genewein, T., Franklin, M., Lerchner, A., Orseau, L., Albanie, S., Bales, A., Wyeth, C., Chan, S., Gabriel, I., Leibo, J. Z., Dafoe, A., Hutter, M., Graepel, T., & Legg, S. From AGI to ASI. Research report, 10 June 2026. arXiv:2606.12683v2 [cs.AI], 30 Aug 2026. https://arxiv.org/abs/2606.12683. Cited pages 2, 27, 28, 37. Authors affiliated with Google DeepMind, University of Waterloo, Australian National University, and University College London. Not official Google policy.
• Executive Order 14148, "Initial Rescissions of Harmful Executive Orders and Actions," signed 20 Jan 2025, 90 FR 8237 (28 Jan 2025); revokes EO 14110 at Sec. 2(ggg). Federal Register: https://www.federalregister.gov/documents/2025/01/28/2025-01901/initial-rescissions-of-harmful-executive-orders-and-actions · White House: https://www.whitehouse.gov/presidential-actions/2025/01/initial-rescissions-of-harmful-executive-orders-and-actions/
• Regulation (EU) 2024/1689 (AI Act), Art. 51 (systemic-risk classification; training compute greater than 10^25 FLOPs presumed high-impact) and Art. 55 (obligations of providers of general-purpose AI models with systemic risk). EUR-Lex: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng · European Commission, "General-Purpose AI Models in the AI Act – Questions & Answers": https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers