ChatGPT crossed 100 million monthly users in January — the fastest consumer product adoption cycle in recorded technology history. Ten weeks after launch, communications professionals across PR, marketing, and corporate communications are already using it. The question is what they should understand about the limitations before they build workflows on top of it.
What follows is the honest read at ten weeks in. The category is moving fast. Some of what is written here will look different in twelve months. Some will not.
The five limitations that matter now
Context understanding is thinner than the output suggests. ChatGPT produces authoritative-sounding text. The confidence of the writing is not a signal of accuracy — it is a feature of the model architecture. On nuanced or specialized subjects, the output can be plausibly worded and factually wrong at the same time. The category calls this "hallucination." It is not a bug the developers will patch out in a point release. It is a structural feature of how these models generate text. Any published output has to be reviewed by a human who actually knows the subject.
The knowledge cutoff is September 2021. ChatGPT is running on GPT-3.5, trained on data through September 2021. Any event, product launch, appointment, or fact that emerged after that date is either absent from the model's knowledge or hallucinated. A press release referencing a company's current executive team, a market update citing recent data, or a competitive positioning piece describing the state of a category cannot be generated accurately without a human filling in what the model does not know. Microsoft's Bing Chat integration, released to preview earlier this month, and Google's Bard, announced a week before Bing, both attach live web retrieval to the model class. That will address part of the cutoff problem for the products that ship the integration. It will not eliminate the underlying hallucination risk.
Tone replication is mechanical. The model produces text that reads as competent generic writing. It does not replicate the specific voice of an executive, a brand, or a house style without heavy prompting and careful editing. The output has a texture — a rhythm, a preference for triads, a tendency toward abstract nouns and hedging — that is now recognizable to editors, journalists, and increasingly to readers. Communications teams that publish AI-generated content unedited are already being called out in trade coverage. The productivity gain is real. The publish-directly gain is not.
Sensitive information handling is an active operational risk. Content submitted to ChatGPT is logged and can be used for model improvement. That has real implications for confidential client information, internal strategy, personnel matters, financial data before disclosure, and privileged legal material. Several large companies — Amazon, JPMorgan Chase, Verizon — have already restricted or banned employee use of ChatGPT for internal work. The New York City public school system blocked ChatGPT on school networks and devices in January. The pattern will continue. Communications teams need clear internal protocols on what can and cannot be entered into the tool. Yesterday.
Scale of deployment inside enterprise workflows is still early. ChatGPT Plus launched February 1 at $20 per month with priority access. The enterprise API is available through OpenAI directly and through Microsoft's Azure OpenAI Service. But the workflow integrations — approval, review, versioning, brand-voice enforcement, compliance logging — are not yet built into most communications operations. Teams using ChatGPT are doing so through the browser interface, one query at a time. Enterprise deployment at the scale that would justify displacing existing workflows is not yet reality.
Two limitations that are emerging faster than the discussion
Prompt injection. Attackers can embed instructions inside documents, web pages, or user inputs that override the operator's instructions to the model. Kevin Roose's conversation with Bing's chat, published in The New York Times today, is one of a growing set of examples of the model being pushed into behavior its operators did not intend. For communications teams, prompt injection matters most in customer-facing chat deployments. Any brand deploying a ChatGPT-class model into a public interface has to plan for adversarial inputs, not just for helpful ones.
Copyright and training data exposure. Getty Images sued Stability AI on February 6, alleging its training set included copyrighted Getty images. Related class actions have been filed against OpenAI-adjacent products. The intellectual property status of AI-generated content, and of the training data behind these models, is unsettled. Communications teams should assume ambiguity for at least the next twelve to twenty-four months, and should be careful about representing AI-generated work as fully cleared.
What communications teams should actually do
Five disciplines that hold up at ten weeks in.
Draft and review, do not publish directly. Use ChatGPT to accelerate drafts, brainstorming, competitive summaries, first cuts at speeches and bylines. Do not publish anything without a human writer who owns the final product. Every piece that ships needs a name behind it that stands over the accuracy, the voice, and the judgment calls.
Write a data-handling policy this week. What can be entered into the tool. What cannot. Who approves exceptions. How the policy is trained into the team. The compliance and legal groups need to see it. The account teams need to work under it.
Verify every fact. Every date, quote, statistic, executive name, product feature, or claim in the output has to be independently verified. The model is confident. The information may not be current or accurate. The verification step is not optional.
Do not put client-confidential material into the browser interface. This is worth restating separately. If the material is subject to an NDA, is embargoed, is pre-announcement, or is legally privileged, it does not go into a browser-based chat interface with terms that allow logging and training use.
Watch the enterprise deployments. Microsoft's OpenAI integration into Word, Excel, PowerPoint, Outlook, and Teams is coming. Google's Workspace equivalents are coming. The productized versions inside enterprise environments will have data-handling terms and integration hooks the consumer browser interface does not. When those ship, the deployment story changes.
The honest read
ChatGPT is a substantial productivity tool for communications work. It is not a substitute for the judgment, voice, and accountability that the discipline still requires. The teams that integrate it thoughtfully — with review, verification, data-handling protocols, and human writers owning the final product — will produce more work, of better quality, at lower cost, in less time.
The teams that treat it as a replacement for junior writers, or as a way to publish without review, will produce hallucinations at scale. Some of them already have. The trade coverage of those cases is only just starting.
The category will look meaningfully different by the end of 2023. The five limitations above are the ones to plan around now.
The Impact of AI on Public Relations · Bing Chat vs. Google Bard: The Race Is On · Getty v. Stability AI and What Comes Next
