Linguistic Engineering: The Communications Profession's Original Mandate Just Returned

The Transition

By the final week of April 2026, the enterprise-agent era had stopped sounding like a roadmap and started behaving like infrastructure. The shift did not begin that week. On February 5, OpenAI introduced Frontier, an enterprise platform for building, deploying, and managing AI agents with shared context, onboarding, feedback, permissions, and boundaries.

On April 22, Google Cloud launched Gemini Enterprise Agent Platform, positioning it as an end-to-end system for agent development, orchestration, and governance. On April 28, Anthropic released Claude connectors for Adobe, Ableton, Autodesk Fusion, Blender, Splice, SketchUp, Resolume, Affinity, and other creative and engineering tools — bringing AI into the software environments where regulated work, creative work, and technical work actually get done. That same day, OpenAI's latest models and Codex became available through Amazon Bedrock, extending agentic capability into AWS's enterprise AI infrastructure.

At the same time, the governance layer was being renegotiated in real time. Through its Digital Omnibus, the European Commission moved to defer the AI Act's high-risk obligations. An initial trilogue on April 28 ended without agreement — but negotiators returned to the table, reached a provisional political agreement on May 7, 2026, and saw it confirmed by member-state representatives in the Council on May 13.The deal postpones stand-alone high-risk (Annex III) obligations to December 2, 2027, and high-risk AI embedded in regulated products (Annex I) — including medical devices — to August 2, 2028, with formal adoption expected ahead of the original August 2, 2026 date.

The stated reason for the delay is itself the argument of this paper. Implementation was outrun by unfinished harmonized standards, unresolved compliance tooling, and undesignated competent authorities — so the obligations were pushed back to wait for the infrastructure to catch up.The capability arrived; the verification layer did not.Regulators bought sixteen additional months because value had outpaced the means to govern it. That is the gap this paper describes, now written into law.

Read separately, these events are corporate and regulatory news. Read together, they are something else. They mark the point at which AI ceased to be a tool one chooses to use and became infrastructure one is now embedded in.

The locus of work — drafting, designing, deciding, coordinating, approving, escalating, and acting — is moving into systems where a human author is no longer the default speaker, where the path from intent to output runs through a probabilistic intermediary, and where the regulatory obligation to govern that intermediary is arriving before the operational infrastructure required to satisfy it.

This is the third structural shift in corporate communications in a hundred years. The previous two — the mass media consolidation of the twentieth century and the internet fragmentation of the early twenty-first — each reshaped the discipline.This one returns the discipline to its original mandate.

The Unobserved History

The discipline of strategic communications was not invented to sell products.It was invented to keep institutions alive when the conditions of their legitimacy changed faster than they could.

On April 20, 1914, the Colorado National Guard, supplemented by Colorado Fuel & Iron Company guards, fired on a striking miners' camp at Ludlow, Colorado. Roughly twenty-five people died, including two women and eleven children who suffocated in a cellar beneath a burning tent. The Rockefeller family, which controlled CF&I, faced not a labor dispute but a legitimacy crisis — and a regulatory environment that could have ended the company. They retained Ivy Ledbetter Lee.

Lee did two things. He instructed the Rockefellers to be visible — to walk the mines, to be photographed shaking hands with miners, to humanize the institution behind the corporate name. And he distributed a Declaration of Principles, originally issued in 1906, committing to provide accurate, attributable information to the press rather than hiding behind silence or paid placement. Lee's Declaration promised the transparent supply of prompt and accurate information to the press and public.

The discipline that emerged from these decisions had a specific function:to reduce the information asymmetry between powerful institutions and the publics whose permission they required to operate.

Three years before Ludlow, Standard Oil had been dissolved into thirty-four entities. The 1911 breakup came through the Supreme Court under the Sherman Antitrust Act — but the antitrust action was made politically possible by Ida Tarbell's serialized investigation in McClure's Magazine (1902-1904), a sustained act of journalistic narrative-building. The lesson institutions absorbed from these decades was not abstract.Coverage shaped legitimacy. Legitimacy shaped survival.Strategic communications, as a profession, was the institutional response to that fact.

The next century saw the discipline mature, and then drift. Edward Bernays, building on his uncle Sigmund Freud's work, argued that consent could be engineered — that publics could be moved through the targeting of subconscious associations rather than the rational presentation of information. His 1928 book Propaganda and his 1947 essay (later expanded into a 1955 book)The Engineering of Consent named the practice. The Torches of Freedom campaign, which he designed for the American Tobacco Company in 1929, recruited women to smoke Lucky Strikes during the New York Easter Parade, repositioning women's smoking as an emancipation gesture.

Bernays inverted Lee's premise. Lee had argued that the institution must reduce information asymmetry to earn legitimacy. Bernays argued that the institution could exploit information asymmetry to manufacture it.

The mass media era that followed operated under Bernays' premises more than Lee's. The discipline organized itself around message control, gatekeeper access, and Advertising Value Equivalency — a metric that measured how much paid media a given earned placement replaced. Trust was assumed, not measured. The audience was a target.

The internet era fragmented the channel landscape and forced the discipline into multi-directional engagement. It also produced the Content Factory: the operational doctrine that competitive advantage came from publishing more, faster, across more surfaces. By the late 2010s, organizations were producing content at a scale no human audience could meaningfully consume. The metric became reach. The substrate became noise.

Then the audience changed. Beginning roughly in 2023, the dominant intermediary between institutions and stakeholders ceased to be only the journalist, analyst, regulator, or search engine result.It became the inference engine— the AI model that reads, summarizes, synthesizes, and answers on behalf of the human at the other end.

By 2026, reported estimates placed Google zero-click search behavior around or above 60%, meaning large portions of the search experience ended without a click to an external website. The user reads the summary and moves on. The institution being summarized has no direct opportunity to correct the synthesis.

A profession built to manage information asymmetry between institutions and human publics now confronts a substrate where the human public increasingly reaches the institution through a probabilistic intermediary. Which is to say:the original mandate has returned, in a new form, with higher stakes.

The Mechanism

The shift is easier to see if one examines what strategic communications has always actually been. Stakeholders never have direct access to an institution. A regulator does not work inside the company. A journalist does not sit in the boardroom. A patient does not read the clinical trial protocol. A retail investor does not audit the books.

Every party who must form a judgment about an organization does so on the basis of a small number of signals — press releases, earnings calls, regulatory filings, conversations with surrogates, third-party coverage, product experience — and infers the institutional reality from those signals.Communications has always been the discipline of engineering those signals.

Not fabricating them — that is the Bernays inversion, and it has been the discipline's recurring failure mode for a hundred years — but selecting them, sequencing them, framing them, and ensuring their consistency across channels and over time. The objective has been to produce, in the inferring mind of the stakeholder, a representation of the institution that is both accurate and aligned with the institution's intent.

What AI has changed is not the existence of inference. It is the speed, scale, opacity, and consequence of inference. A regulator who once read a clinical trial summary may now receive a synthesized answer from a model that has read that summary, three competing summaries, twelve adjacent academic papers, and the comment section of an industry forum.

The synthesis happens in milliseconds. The provenance of any single claim in the synthesis may be difficult to trace. The stakeholder may not be aware that synthesis occurred. And the stakeholder may act on the synthesis as if it were direct knowledge.

This is what makes the current moment structurally different from the prior two shifts. Mass media changed the channels through which signals reached stakeholders. The internet changed the volume and bidirectionality of those signals.AI changes the substrate of inference itself.The intermediary between the institution and the inferring stakeholder is now a system, not a person. The system's representation of the institution is what the stakeholder treats as the institution.

That alone would be enough to require a new communications doctrine. But the risk compounds as AI becomes agentic. An AI-mediated output no longer merely describes. It can route, request, recommend, negotiate, escalate, approve, and act. A hallucinated premise can become a wrong instruction. A wrong instruction can become a permission error. A permission error can become an agent-to-agent handoff that moves across systems before any human sees the drift.

The risk is no longer just a bad summary. It is the compounding security vulnerability of meaning drift inside operational systems. Conventional security can determine what an agent is allowed to access.It cannot, by itself, determine whether the meaning the agent carries remains authorized.Access control can say whether a system may touch a file, invoke a tool, or call another agent. It cannot fully answer whether the agent is still preserving the source, claim, evidence, permission, audience, and intent it was supposed to preserve.

That is the new control gap. Communications has always been about controlling how institutions are inferred. The current moment makes the inference machinery visible, instrumentable, and — for the first time — directly addressable.The discipline does not need a new purpose. It needs a new substrate to operate on. The substrate is the model.

The Trust Evolution

Every transformative technology passes through three phases of trust.

In the first phase, trust is explicit. The technology is visible. Its use is deliberate. People ask whether to use it, what it can do, where it should not be used. The conversation is conscious. Skeptics are vocal. Boundaries are negotiated in public.

In the second phase, trust is implicit. The technology has matured enough that asking whether to use it is no longer the dominant question. The question becomes how to use it well. The conversation moves from category-level skepticism to use-case calibration. The technology is not yet invisible, but it is no longer foregrounded.

In the third phase, trust is systemic. The technology has become infrastructure. People do not consciously decide to trust it; they encounter the world through it. Skepticism does not disappear, but it migrates from the technology itself to specific deployments or specific actors using it. The default has flipped from "why would I trust this" to "why would I question this."

The historical answer to that question is rarely encouraging until enforcement makes it so. Electricity moved through a roughly forty-year arc from novelty to infrastructure between the 1880s and the 1920s. The trust came to be earned only because the institutional substrate caught up. Underwriters Laboratories was founded in 1894 to test electrical equipment to standards that insurers could underwrite. The National Electrical Code, first published in 1897, gave municipalities a basis for permit issuance and inspection.

The infrastructure that made electricity safe enough to disappear into the wall was not the technology itself. It was the testing regime, the code, the inspection authority, and the legal liability framework that enforced compliance. Systemic trust in electricity is justified because the infrastructure that earns it exists.

Aviation took roughly fifty years to follow a similar path, ending with the Civil Aeronautics Board, the Federal Aviation Administration, and the International Civil Aviation Organization establishing the multilayered safety regime that makes commercial flight, statistically, the safest mode of transportation in human history.

Pharmaceuticals followed a shorter and more painful path. The 1962 Kefauver-Harris Amendment, passed in the wake of the thalidomide tragedy, required manufacturers to demonstrate efficacy and safety before marketing — establishing the modern claims-evidence regime that distinguishes pharmaceutical assertions from other commercial speech.The systemic trust that patients now place in prescribed medication is justified because, and only because, the infrastructure of clinical trial design, regulatory review, post-market surveillance, and labeling discipline exists.

In each case, the technology entered systemic trust after the infrastructure that earned the right to that trust was built. The infrastructure preceded the invisibility. Where the infrastructure did not precede it — early aviation before federal oversight, early pharmaceuticals before Kefauver-Harris, early electrical installations before code adoption — the cost was paid in casualties before the regulatory system was forced into existence.

AI is exiting its visible phase faster than any of these technologies. The late-April 2026 sequence is not merely a marketing moment. It is a visible acceleration point in AI's transition from explicit tool to operational substrate.The question of whether systemic trust in AI is earned or complacent is, at the moment, open.

Two weeks after this paper was first published, that question stopped being hypothetical. On May 19, 2026, at Google I/O, Google made a Gemini-powered AI Mode the global default for Search — what it called its largest overhaul in twenty-five years. AI Mode had already passed one billion monthly users, with queries more than doubling every quarter; Google extended it, through Personal Intelligence, to nearly 200 countries and territories in 98 languages, and began retiring the search box itself in favor of conversational, multimodal, agentic answers. The trade press called it the day search stopped being a search engine.

This is the harbinger. A transformative technology earns its way into the background exactly as every prior one did — by delivering enough value, often enough, that the user’s hesitation to trust a system they cannot inspect is quietly overcome. People did not adopt the inference layer because they understood it; they adopted it because the answer was good, and then the asking-whether stopped. When a billion people reach the world through an AI answer by default, the explicit phase is ending in public — the technology is sliding from foreground tool to background infrastructure in real time.

The background transition is being driven by value, not by verification. AI is becoming invisible because it is useful — not because the infrastructure that makes its invisibility safe has been built.

That is the precise danger this paper names. The march to systemic trust is being completed by consumer utility years ahead of the testing regimes, claims-evidence discipline, provenance, and audit infrastructure that earned electricity, aviation, and pharmaceuticals their place in the wall. Value is retiring the skeptics before verification has done its work —premature invisibility at global scale. And it sharpens the mandate rather than softening it: the moment AI Mode became the default front door to the world’s institutions, every institution’s representation became something an inference engine renders — for a billion people, in languages and markets the institution never reviewed. The Algorithmic License to Operate is no longer a regulated-industry edge case; it is the precondition for being understood at all. Closing that gap — building the verification layer beneath the invisibility while the invisibility is still arriving — is the work Signal Fidelity Group was formed to do.

The Mandate

The infrastructure required is not, primarily, a technology problem. It is a representation problem.

What an institution looks like to an inference engine — what claims it makes, what evidence supports those claims, how those claims are versioned and updated, where those claims appear in machine-readable form, how the institution's voice is distinguished from synthesized approximations of it — is a question of how organizational reality is encoded for probabilistic systems.

That work is closer in nature to what strategic communications has done for a century than to what software engineering does.It is the discipline of engineering signals so that inference produces accurate representation.It is the discipline that owns the institutional voice. It is the discipline that mediates between institutional reality and stakeholder understanding.

The economic consequence is now becoming visible. Organizations do not realize AI ROI because models generate more output.Output is activity.Value begins only when that output can be approved, deployed, reused, monetized, audited, and defended. A model that produces a thousand drafts no one can approve has not created leverage. It has created review debt. A system that generates summaries no one can trace has not accelerated the business. It has increased the cost of verification.

In low-stakes environments, the cost of unverified AI output appears as inefficiency. In high-stakes environments, it becomes liability. Trust, in regulated communications, is not an ethics sidebar — it is the economic conversion layer between AI activity and business value.

This is most obvious in regulated industries. In pharmaceuticals, medical technology, financial services, and other claims-governed environments, an AI hallucination is not a software defect in the ordinary sense. It is an unauthorized representation. It can compromise a submission, contaminate a review process, misstate evidence, create downstream regulatory exposure, or halt commercialization until the institution can prove what was said, why it was said, where it came from, and whether it was approved.

The risk is not merely that the machine is wrong.The risk is that the organization cannot demonstrate control over the meaning the machine produced on its behalf.Trust, in this environment, is not an ethics sidebar. It is the economic conversion layer between AI activity and business value.

The same problem compounds as AI systems become agentic. A generated sentence can become an instruction. An instruction can become a workflow. A workflow can become a handoff. A handoff can become an autonomous action taken in the name of the institution. Once that happens, the communications problem becomes inseparable from the governance problem. An institution must be able to prove not only that an agent had permission to act, but that the action preserved authorized meaning.

That is why communications can no longer be treated as a downstream content function. What the discipline has not been, in its drift through the mass media and internet eras, is a discipline that thought of itself this way. Communications professionals were trained as storytellers, message-makers, channel managers, and reputation custodians. The work of engineering the substrate on which inference happens has not been part of the discipline's self-conception.It needs to become so.

The pivot is not from communications to engineering.It is from communications-as-storytelling to communications-as-architecture.

The artifacts the discipline produces — press releases, executive remarks, product communications, regulatory filings, internal narratives, brand assets, claims libraries, approval packets, evidence tables, response guides — are no longer merely texts. They are inputs to inference systems. The systems read them, weight them, retrieve them, synthesize them, and produce outputs that stakeholders then encounter. Whether those outputs are accurate is not, primarily, a question of writing skill. It is a question of whether the inputs were structured for the systems that will read them.

A second reframing helps. The discipline's purpose, in what we now call the social license to operate, was to earn the public permission that legal authority alone could not provide. The infrastructure of that license was a relationship between institutions and the publics that granted permission. In an era where AI systems mediate between institutions and publics, the social license is downstream of an algorithmic license.

If the inference engines that summarize the institution to its stakeholders cannot produce accurate, attributable, verifiable summaries, the institution loses access to the conversation in which its social license is granted.The Algorithmic License to Operate is the new precondition. It is upstream of every other communications objective.The discipline that built the social license is the discipline equipped to architect the algorithmic one. The reason is not nostalgia. It is fit.

The Principles

Three principles follow from the analysis. They are starting positions, not endpoints. The discipline of Linguistic Engineering— the term for this work — will mature them.

Principle 01 — Trust must be earned before it can be assumed.

The path from explicit to systemic trust runs through infrastructure. There is no shortcut through marketing, policy declaration, or category-level reassurance. An institution that wants its AI-mediated representation to be trustworthy must build the systems that make trustworthiness verifiable: structured claims, attributable provenance, drift instrumentation, audit trails, intervention records, and repeatable approval logic.

This is where the term Linguistic Engineering must be understood mechanically, not metaphorically.Linguistic Engineering treats institutional language as structured data with enforceable rules, traceable provenance, and verifiable fidelity.Claims become objects. Evidence becomes linked infrastructure. Approvals become state changes. Permissions become constraints. Audiences, jurisdictions, versions, and use contexts become machine-readable conditions.

Outputs are not simply generated; they are evaluated against the reference signals they were authorized to preserve. This complements probabilistic prompting with deterministic structural constraints. A prompt asks the model to behave. A linguistic control system defines the conditions under which the output is allowed to stand.

In practice, this means claims registries, evidence maps, approved-language objects, provenance ledgers, drift scores, remediation paths, escalation thresholds, and audit packets. Closed-loop control techniques operating around the model — comparing outputs against authorized references, correcting where possible, blocking or escalating where required — function as the operational layer that makes meaning enforceable rather than aspirational.The model remains probabilistic. The system around it does not have to be.

That work is unglamorous. It is also what distinguishes earned trust from complacency. The late-April 2026 acceleration will not, on its own, force this work. Enforcement, when it arrives, will.The discipline that begins the work before enforcement arrives sets the standard. The discipline that waits for enforcement inherits one.

Principle 02 — Every AI-mediated output is an intermediary, not a final speaker.

A press release written with AI assistance, a product description summarized by a retrieval system, an executive answer reconstructed from a quarterly transcript, a regulatory disclosure read through an enterprise model — none of these is a direct utterance of the institution. Each is the output of a probabilistic system operating on institutional inputs.

The discipline that governs intermediaries — claims-evidence pairing, provenance disclosure, version control, semantic consistency across channels — applies to AI-mediated outputs by default. The institutional voice is no longer expressed only through human-authored text.It is expressed through the systems that synthesize on the institution's behalf, whether the institution acknowledges those systems or not.

The same principle applies to agentic systems. An AI agent acting for an institution is an intermediary with operational consequences. It carries identity, intent, permission, scope, evidence, escalation, and handoff obligations. When an agent engages another agent, the institution has not escaped communications governance. It has intensified it. The question is no longer only, "Did we say the right thing?" The question is,"Did every system acting on our behalf preserve the meaning it was authorized to carry?"

Principle 03 — The infrastructure that makes a technology invisible is what determines whether the invisibility is safe.

This is the operational form of the trust evolution claim. Electricity is safe to forget about because testing regimes, electrical codes, inspection authorities, and liability frameworks exist. Aviation is safe to forget about because certification, investigation, air traffic control, maintenance discipline, and international standards exist. Pharmaceuticals are safe to trust because clinical trial design, regulatory review, labeling discipline, post-market surveillance, and enforcement exist.

AI is in the process of becoming invisible. The question is whether the infrastructure that earns the right to that invisibility — the equivalent of testing regimes, certification bodies, claims-evidence discipline, auditability, inspection, and enforcement — is being built fast enough. Where it is not, the disappearance is premature.Premature invisibility is the failure mode the next decade will pay for.

The Close

There is a personal disclosure that belongs here. I have spent the past two decades inside three of the most regulated communication environments in the United States: Johnson & Johnson MedTech, Takeda, and Boston Scientific.The discipline I am describing is not abstract to me.It is the discipline I learned to practice when claims required clinical evidence, when product communications required approved language, when regulatory filings required exact provenance, and when the cost of representation drift was patient harm.

The current moment is not, for that reason, a category-level disruption.It is the realization that the discipline of regulated communications — the discipline that has always operated under explicit constraint, instrumented evidence, and traceable claims — is the model the rest of the field will need to adopt.

In March 2026, I formed Signal Fidelity Group to build the infrastructure that operationalizes that discipline.Linguistic Engineering is the substrate. LIRA — Looped Inference Regulation Architecture — is the control architecture.

The body of work is integrated. The thesis is single.Human authorship. Organizational meaning. Agent-to-agent action. Three surfaces. One control problem.

What follows in the coming months is a deeper accounting. The historical arc, in detail. The science of engineering trust through language. The empirical foundation behind the thesis. The operational handbook for practitioners. Each is a chapter of the same argument.

The communications profession's mandate has not changed since Ivy Lee articulated it in 1906. The substrate it operates on has. The discipline that engineered the social license is the discipline that will engineer the algorithmic one.We are not at the end of strategic communications. We are at the beginning of its highest-stakes chapter.

Abhi Basu is the founder of Signal Fidelity Group, the trust infrastructure layer for AI. He spent twenty years leading strategic communications at Johnson & Johnson MedTech, Takeda, and Boston Scientific. He writes about Linguistic Engineering at signalfidelitygroup.com.