Most AI governance regimes start from the opposite premise. The EU's AI Act sorts systems into risk tiers and requires conformity assessments before high-risk deployment; a company must show its work before it ships. Israel's National AI Directorate, in the five-year National AI Strategic Plan it released on August 13, 2026 to implement Government Resolution 4255 ("Accelerating Artificial Intelligence in Israel and Establishing Global Leadership," approved by the cabinet on June 16, 2026), proposes to run the presumption the other way. Under the plan's "right of way for AI" principle, an action an AI system can perform better than a human should not be blocked by regulation unless regulators first prove a "specific and material risk." Public comment on the plan closes September 6, 2026.
The steelman for precaution
The case for the EU's sequencing is not frivolous. High-risk automated decisions — credit, hiring, medical triage, policing — can cause harm that is diffuse, hard to trace back to a specific model, and expensive to unwind after the fact. Requiring evidence of safety before deployment, rather than after, is the same logic behind drug approval and financial-product suitability rules: the burden sits on the party with the information advantage and the profit motive, not on the diffuse public that would bear the cost of getting it wrong. Regulators who have watched algorithmic harm compound quietly for years before anyone could prove causation have reason to distrust a system that asks them to wait for the smoking gun.
Why Israel is betting the other way
Israel's directorate has direct evidence of what the precautionary alternative costs. Reporting from Calcalist's CTech describes Israeli AI firms — including the content-safety developer Netspark — citing the EU AI Act's compliance costs and jurisdiction-by-jurisdiction ambiguity as a reason to redirect growth and R&D toward the US and UAE rather than Europe. That is the fear the "right of way" principle is written against: a regime where the cost of proving a system is not dangerous falls hardest on the small companies that can least afford to carry it, before any actual harm has been shown to exist. For a country whose AI sector is disproportionately startups rather than incumbents, shifting the burden of proof onto regulators — who must show specific and material risk, not hypothetical risk — is a defensible bet if state capacity can back it up.
That capacity is the open question. The plan pairs the deregulatory principle with an "Israeli Safety Card," a domestic autonomous-systems safety standard the Directorate wants EU and US regulators to mutually recognize — in effect, asking Brussels and Washington to treat an Israeli sign-off as equivalent to their own. Mutual recognition only works if the certifying body has the technical depth and independence to make the sign-off credible, and that is where the plan's own numbers cut against it. Calcalist's CTech reported the Directorate's 2026 budget at roughly NIS 120 million with 20 staff positions, rising to NIS 13 million a year in 2027 for salaries and operations — a fraction of what US and EU AI-safety institutes operate with. Globes' Hebrew-language business desk was blunter, quoting an industry source calling the broader plan "a drop in the ocean" against a total 2026 government AI allocation of roughly NIS 3 billion, some of it still subject to cuts.
The governance gap that undercuts the pitch
There is also a structural tension the plan does not resolve. Brig. Gen. (res.) Erez Askal, who heads the Directorate, was appointed without a competitive tender — the Civil Service Commission approved the exemption citing urgency, according to Calcalist's CTech — and the Directorate sits inside the Prime Minister's Office rather than as an arm's-length regulator. A "right of way unless proven harmful" standard depends entirely on regulators willing and able to actually prove harm and act on it; housing that function inside the political center of government, with a small budget and a director installed by fiat rather than process, is precisely the setup that makes critics of light-touch regimes nervous. It does not disprove the underlying principle, but it means the principle's credibility now rests on an institution that has not yet demonstrated it can do the proving.
What to watch
The strongest version of Israel's bet is right on the economics: undifferentiated precaution taxes small AI companies more than it protects anyone, and a narrower "prove the specific harm" standard is a more honest way to write that trade-off into law than pretending risk tiers are self-evident. But a proof-of-harm regime is only as good as the body doing the proving, and right now that body is under-resourced relative to its own ambitions and was staffed outside the normal process meant to establish its independence. The September 6 consultation is the moment to press for two things the current draft is thin on: how "specific and material risk" will actually be adjudicated case by case, and whether the Directorate's budget and staffing will scale to match the certification role it is asking the EU and US to trust. Absent that, "right of way" risks becoming a slogan a small office cannot back up rather than the credible alternative to the AI Act it is pitched as.