The Legal AI 'Change in Law' Clause Problem: Why Clause Detection Is Not Risk Analysis
There is a performance gap in AI contract review that nobody in the legal technology industry is eager to discuss openly. AI tools—including the category leaders—are getting considerably better at finding clauses. They are not getting meaningfully better at understanding what those clauses actually do....
There is a performance gap in AI contract review that nobody in the legal technology industry is eager to discuss openly. AI tools—including the category leaders—are getting considerably better at finding clauses. They are not getting meaningfully better at understanding what those clauses actually do. Nowhere is this gap more commercially consequential than in the review of change in law provisions, where a surface-level summary can be worse than useless: it creates the illusion of diligence without the substance of it.
What the Tools Actually Do
Ask any leading AI contract review platform to assess a change in law clause and you will generally receive a coherent summary. The tool will identify the clause, describe it as allocating regulatory risk between the parties, perhaps note whether the adjustment mechanism favours the supplier or the buyer, and flag it as reviewed. In a playbook-driven workflow, it may even compare the clause against a fallback position and report a pass or a deviation.
What you will not receive—at least not reliably, and not without significant further prompting—is an answer to the questions that actually determine commercial exposure. Does the definition of "law" in this contract extend to subordinate legislation, regulatory guidance, and administrative directions, or does it capture only primary statutes? Does a triggering change give the affected party a right to renegotiate, a right to terminate, or merely a right to notify? Is there a materiality threshold, and if so, how is it calculated? These are not edge questions. They are the operative questions. The clause summary is not the analysis.
Why This Matters in Infrastructure and Power Purchase Agreements
Consider a long-term power purchase agreement in a jurisdiction where electricity regulation is evolving rapidly—a common situation across European and Asia-Pacific markets. A change in law clause that refers only to "applicable laws and regulations" enacted by the legislature may not capture a grid operator's updated technical connection standards, an energy regulator's revised dispatch instructions, or a government policy direction that restructures ancillary service markets. These instruments—typically subordinate legislation, delegated acts, or regulatory guidance—are the instruments that actually change the economics of generation assets mid-contract.
If the definition of "law" in the PPA does not include these instruments, the change in law clause provides no relief when they move. A tool that identifies the clause and reports that regulatory risk is allocated has told the GC something true and something useless. The commercially decisive question is whether the clause covers the specific regulatory instrument that is about to change the project's revenue stack. That requires reading the definition section, cross-referencing it against the jurisdiction's regulatory architecture, and forming a judgment about what is and is not captured. Current AI tools will occasionally surface the definition, but they do not reliably connect it to the coverage question.
The Trigger Problem in Pharmaceutical Licensing
Pharmaceutical licensing agreements present a different but equally instructive version of this failure. A change in law clause in a territory licence will often specify that a regulatory change affecting the commercialisation of the licensed product entitles one party to adjustment. The adjustment mechanism—whether it operates as a royalty reduction, a renegotiation obligation, or a termination right—determines entirely whether the clause provides meaningful protection or merely documents a grievance.
AI tools reviewing these agreements routinely identify the change in law clause and characterise it as a risk-allocation mechanism without distinguishing between these outcomes. A renegotiation obligation with no defined process, no timeline, and no fallback is not materially different from having no clause at all if the parties cannot agree. A termination right triggered only by changes that render commercialisation "impossible" imposes a threshold that regulatory guidance changes—as opposed to outright prohibitions—will almost never satisfy. The practical effect of the clause depends entirely on trigger language and mechanism design. Flagging the clause's existence does not address either.
Regulatory Guidance: The Definitional Gap That Costs the Most
The category of instrument most consistently missed by AI clause analysis is regulatory guidance—agency-issued positions, enforcement priorities, interpretive letters, and informal directions that do not have statutory force but that, in practice, determine whether a product can be sold, a project can proceed, or a contract can be performed. In the pharmaceutical context, this includes agency guidance on manufacturing standards, labelling requirements, and post-approval study obligations. In infrastructure, it includes planning policy statements, environmental agency positions, and sector-specific regulatory frameworks that operate beneath the level of legislation.
Whether a change in law clause captures these instruments is almost entirely a function of how "law" is defined. A definition that reads "any statute, regulation, rule, or order" excludes regulatory guidance as a matter of plain language. A definition that adds "any directive, standard, or requirement of any governmental or regulatory authority" may or may not capture it depending on whether the relevant guidance is characterised as a "requirement." This is a judgment call that depends on the specific language, the jurisdiction, and sometimes the regulatory body's own published position on the legal status of its guidance.
An AI tool that summarises a change in law clause without parsing the definition of "law" and connecting it to the instrument type the client actually faces has not performed contract review. It has performed contract indexing.
The Standard You Should Be Holding AI Tools To
The clause summary is the beginning of the analysis, not the end of it. Transactional lawyers and GCs using AI contract review tools for change in law provisions should be demanding outputs that address at minimum: the scope of the "law" definition and whether it captures subordinate legislation and regulatory guidance; the trigger threshold and whether it is objective or discretionary; the remedy mechanism and whether it provides renegotiation, price adjustment, or termination; and any materiality filters and how they interact with guidance-level changes.
If the tool you are using does not produce that analysis without extensive manual prompting, it is a drafting assistant that has been sold as a risk assessment engine. The distinction matters. In infrastructure and pharmaceutical contracts that run for decades across shifting regulatory landscapes, mistaking clause detection for risk analysis is not a workflow inefficiency. It is a liability.