A plain-language primer for elected officials and the people who advise them. No jargon and no sales pitch. Products are named only in Appendix B, as a neutral reference list.
When people say "AI" today they almost always mean one specific thing: a system that reads a request written in plain language and produces a plain-language response. It can draft a letter, summarize a report, answer a question, translate a document, or write a line of computer code. This kind of system is called a large language model, and the products built on it are usually called assistants or chatbots. That is the technology reshaping offices, and it is the subject of this briefing.
It is worth being equally clear about what this technology is not. It is not a robot, and it is not a mind. It does not know things the way a person knows them, it is not connected to a live database of facts unless someone deliberately connects it to one, and it is not conscious, no matter how natural the conversation feels. It is a very sophisticated pattern-completion engine. Understanding that one idea is most of what a decision-maker needs.
The reason it seems to understand is that it was trained on a large share of the writing humans have published, so the patterns it learned are rich enough to hold a coherent conversation, keep track of context, and adjust its tone. When you ask it to write a memo in a formal voice, it produces formal-sounding text because it has seen a great deal of formal writing and learned what that pattern looks like. The result can be genuinely useful. It can also be confidently wrong, because predicting fitting words and stating true facts are not the same thing. That distinction runs through this entire briefing.
The underlying idea is decades old, but three things changed at once in the last few years: the models became far more capable, they became cheap enough to offer to the public, and they became easy enough to use that no technical training is required. That combination is why the technology moved from research labs into everyday office software in a very short time, and why questions about its use, its cost, and its risks are now landing in front of public bodies rather than only technology departments.
Four steps take you from a pile of text to an assistant that answers questions. Each step also explains one of its weaknesses.
Because it learned from a fixed body of text gathered up to a certain date, it does not know what happened after that date unless it is separately connected to live information. Ask a plain assistant about last week and it may answer from patterns rather than facts, or state that it cannot know.
Because it holds patterns rather than a database, it can produce a fact that sounds right, is phrased with total confidence, and is simply false. This is common enough to have a name, covered in the limits section.
Because it predicts fitting words, it reflects whatever was common in its training text, including the biases in that text. It is not neutral by default; it is average by default, and average includes society's blind spots.
Because it was tuned to be agreeable, it will often go along with a flawed premise in your question rather than challenge it. It has to be asked, directly, to push back or to show its uncertainty.
Not all assistants do the same job. They fall into a handful of families, and the family matters more than the brand.
Products come and go and rename themselves constantly, so it is more durable to think in families. Most tools a public body will encounter fall into one of these, and some products belong to more than one.
The family decides what the tool touches and where your information goes, which is what a public body actually needs to govern. A general assistant answers from what it learned and touches nothing of yours. An answer engine reaches out to the live web. A suite-embedded assistant reaches into your own mail and files. A task agent takes actions on your behalf. A self-hosted model keeps everything inside your walls. Two tools can look identical in the chat window and carry completely different risk depending on which family they belong to.
Beneath the branding, tools differ on a few features that decide whether one is right for public-sector work.
Some tools do their processing on the provider's servers in a distant data center, which means everything you type is sent out of your building to be handled by another company. Others can run partly or entirely on your own device or inside your own network, so that sensitive material never leaves your control. A tool that processes locally is meaningfully stronger on privacy, because information that never travels cannot be intercepted, retained, or subpoenaed from a third party. For any work touching resident records or personnel files, where the processing happens is a first-order question, not a technical footnote.
Some tools answer only from the patterns learned during training, and give you no way to check where a claim came from. Others retrieve live source documents first and attach a citation to each fact, so a reader can click through and verify. For public work, where a statement may end up in a staff report, a council packet, or a response to a resident, a tool that cites its sources is far safer than one that asks you to take its word. The presence of citations does not guarantee accuracy, but their absence guarantees you cannot check.
This is the feature most likely to disqualify a tool for public use, and it has nothing to do with quality or price. Some services store and process everything you send them on servers located in foreign countries, under legal systems where that country's government can compel access to the data. A tool may be capable, inexpensive, and popular, and still be unacceptable for a public body because the resident information you feed it could be reached by a foreign state with no recourse available to you. Governments have already acted on this. For example, on Jan 31, 2025, the Texas Governor barred several foreign-developed apps, including an AI assistant, from state-issued devices and networks.[3] The question to ask is simple: in which country does our data physically live, and whose laws govern who can see it?
On many consumer plans, everything typed into the tool may be used by the provider to further train its models, which means your text can influence future answers given to strangers. On the business and government plans of the same products, this is usually turned off by written commitment, and the data is walled off. The default on a free account is often the least protective setting. For any official use, the plan and its data terms matter as much as the tool.
Some tools first retrieve relevant documents, such as your own policies or the live web, and then write the answer from them. This is called retrieval-augmented generation, or RAG. It makes answers easier to check, but the answer is only as good as the documents it retrieved.
Some models are "open weight," meaning the model itself can be downloaded and run by anyone, including on your own hardware. Others are "closed," offered only as a service you connect to. Open models give an organization the option of full control and self-hosting; closed models are simpler to adopt but keep you dependent on the provider. Neither is automatically better, but the choice affects both privacy and long-term cost.
| Feature | The safer answer for a public body |
|---|---|
| Where processing happens | On your own device or network where possible; otherwise a provider with data centers in your own country. |
| Sources and citations | Cites where each fact came from, so staff can verify before it goes in a public document. |
| Data location and jurisdiction | Data stored under your own country's law, never on servers a foreign government can compel. |
| Training on your inputs | A written commitment not to train on your data. Usually the business or government plan, not the free one. |
| Open or closed | Decided by your privacy needs and staff capacity; open enables self-hosting, closed is simpler. |
Constituents ask whether AI is straining the grid and the water supply. The honest answer is that it depends on the data center, and the newest ones are a different animal from their reputation.
The assistants in this primer do their heavy work in data centers, large buildings full of computers. Training a model and answering millions of requests draws electricity, and keeping the computers from overheating has traditionally taken a great deal of water. The headlines about AI's footprint, and any proposal to build a data center near your community, are about these buildings. The concern is real and worth understanding, because the picture is changing quickly and the reputation is a generation out of date.
First-generation data centers were genuinely heavy on local resources. They cooled their computers by blowing chilled air through large rooms and by evaporating water, which strained local water supplies and electrical grids. Lawrence Berkeley National Laboratory estimates that U.S. data centers used about 66 billion liters (about 17 billion gallons) of water directly in 2023, mostly for cooling (as of Dec 2024).[4] Much of the power they drew was spent on that cooling rather than on computing itself. When people picture AI draining a reservoir or overloading a grid, they are picturing this older design, and they are not wrong about it.
Two things are true at once, and it is important not to let one answer for the other. Each unit of AI work is getting dramatically more efficient. At the same time, the total amount of AI work is rising so fast that overall electricity demand from data centers is still climbing. U.S. data centers used about 4.4% of the nation’s electricity in 2023, and that could reach 6.7% to 12% by 2028 (as of Dec 2024).[4] Worldwide, the International Energy Agency estimates data centers used about 415 TWh in 2024, or about 1.5% of global electricity, and projects about 945 TWh by 2030 (as of Apr 2025).[8] Efficiency per task and total consumption are separate questions. A vendor who answers a question about total demand by pointing to per-task efficiency is changing the subject, and so is a critic who answers a question about a modern facility by describing an old one.
This is where the topic stops being abstract for an elected official. If an operator comes to your community, these are the questions that separate a good neighbor from a resource drain, and a credible operator will answer all of them in writing.
For the history of data center cooling, see The data center comes full circle.
A neutral summary of laws and policies that already apply. This is not legal advice; confirm details with counsel.
The EU AI Act entered into force Aug 1, 2024. Bans on certain practices and AI-literacy duties have applied since Feb 2, 2025. Rules for general-purpose AI (GPAI) models have applied since Aug 2, 2025. Most other rules, including transparency duties, apply from Aug 2, 2026. An amendment (the “Digital Omnibus on AI,” in force Jul 27, 2026) moved the rules for high-risk uses to Dec 2, 2027, or Aug 2, 2028 for AI built into regulated products. It also adds bans on AI that generates non-consensual intimate images or child sexual abuse material, which apply from Dec 2, 2026. It matters locally when a vendor serves EU customers, because its products may be built to these rules.[21]
Status as of Oct 26, 2026. Laws in this area change often; check the linked sources before relying on any item.
The failures are predictable, which means they can be managed. The danger is not that it breaks loudly, but that it fails quietly and confidently.
The best-known failure has a name: a "hallucination." The tool will state a figure, a citation, a court case, or a quotation that does not exist, phrased with the same confidence as a true one. It is not lying, because it has no concept of truth; it is completing a pattern. Any fact that will be relied upon must be checked against a real source.
A plain assistant knows nothing after the date its training data was collected. It may not tell you that, and may answer a question about a recent event by guessing. Only a tool deliberately connected to live sources can speak to the present.
It learned from human writing and mirrors the imbalances in that writing. Left unmanaged, it can carry those imbalances into hiring language, enforcement, or benefit decisions, which is why high-stakes uses need human review.
It is very good at appearing to reason, and it can work through many problems correctly. But it can also fail basic logic or arithmetic while sounding authoritative, because it is predicting the shape of an answer, not computing one. Do not trust an unchecked number.
Because it was tuned to be helpful, it tends to accept the premise of your question and tell you what you seem to want to hear. Ask it to argue the other side, list what it assumed, or say what it is unsure of, and it will, but only if asked.
A person, not the tool, is responsible for the result. The technology can draft, summarize, translate, and speed up nearly any written task, and it will save real time doing so. What it cannot do is carry accountability. When an assistant drafts a notice, a policy, or a public statement, the public official who issues it owns every word, including the ones the tool got wrong. That principle, more than any technical control, is what keeps its use safe.
Where it helps, where to be careful, and the questions to ask any vendor before a tool touches public information.
Never put confidential resident data, personnel records, or anything not already public into a consumer-grade tool. Once it leaves your control you cannot get it back, and on many free plans it may be used to train the provider's model.
Remember public records law. A prompt and its answer may themselves be public records subject to disclosure and retention. Decide how they are logged before staff start typing, not after.
Keep a person accountable for any output that reaches a resident, a council packet, or the public. The tool drafts; a named human approves and owns it.
Be transparent. Residents increasingly expect to be told when they are reading AI-assisted material or speaking to an automated assistant. A short disclosure policy prevents a larger problem later.
These six questions, answered in writing by the vendor, separate a tool that is safe for public use from one that is not. Every one of them traces back to a feature covered in this briefing.
Train staff before giving them tools; in Texas, local employees and officials who use a computer for at least 25% of their duties must already complete DIR-certified AI training each year.[2] Adopt a short acceptable-use policy, involve employees and any bargaining units early, and treat AI as a change in how jobs are done, with written expectations about review and quality.
When AI helps make decisions about people, such as hiring, benefits, permits, or enforcement, keep a record of what the tool did, review it for bias, and give residents a way to reach a human and appeal. Liability generally stays with the body that made the decision, not the tool.
The bottom line for a public body. This technology is genuinely useful and is not going away, and a blanket ban tends to push staff toward unmanaged consumer tools rather than eliminate the risk. The workable path is the opposite: choose tools whose data stays under your own country's law, that cite their sources, and that commit in writing not to train on what you give them; keep confidential information out of anything consumer-grade; and keep a named person accountable for every output that reaches the public. Get those three things right and the rest is detail.
The right tool is rarely the cheapest or the most talked-about. It is the one whose answers to the six questions above you can live with.
This briefing is educational and vendor-neutral. The briefings name no products by design, so they stay accurate as tools change. It is not legal advice; a public body should confirm its own obligations under applicable public-records, privacy, and procurement law with counsel.
XOtavo advises organizations on the technology behind their operations, including how to adopt AI tools safely. Prepared by XOtavo. Questions: help@xotavo.com or 713.980.2000. xotavo.com
Synthetic voices and video are now easy to make. A short plan, made in advance, matters more than any detection tool.
A “deepfake” is synthetic audio, video, or images made to look like a real person saying or doing something they did not. A short sample of someone’s voice can be enough to imitate it, so officials and candidates are likely targets.
In Texas, publishing a deceptive deep fake video within 30 days of an election, with intent to injure a candidate or influence the result, is a crime.[14]
AI can now produce realistic sexual images of real children, and some chatbots are designed as “companions” that young people talk to for hours. Texas law covers obscene AI-generated images that appear to depict children, and federal law requires platforms to remove non-consensual intimate images within 48 hours of a valid request.[15,16,17] Questions for schools and libraries: Which AI tools can students use? What do they collect? Who reviews incidents, and how are families told?
Assistants that talk in natural, lifelike voices and remember earlier conversations are arriving quickly. For a public body, the same rules hold: tell residents when the voice is not a person (in Texas, a legal duty for agencies), govern stored conversations as records, and treat lifelike voices as a fraud-awareness matter.[1]
Every term of art in this primer, in one plain sentence each. Keep it in front of you in the meeting.
Terms are defined for a general audience and kept deliberately brief. Where a precise legal or technical definition matters, confirm it in context.
XOtavo advises organizations on the technology behind their operations, including how to adopt AI tools safely. Prepared by XOtavo. Questions: help@xotavo.com or 713.980.2000. xotavo.com
The briefings name no products on purpose. This appendix is a neutral, dated reference list of well-known names, grouped by family.
This list puts names to the families described in Briefing 3. It is a snapshot, not a ranking or a recommendation, and inclusion means only that the product is widely known. Many products belong to more than one family. Whatever the name, the test is the same: the six questions in Briefing 7.
| Family | Examples (maker) |
|---|---|
| General assistants | ChatGPT (OpenAI) · Claude (Anthropic) · Copilot (Microsoft) · Gemini (Google) · Grok (xAI) |
| Answer engines | Perplexity (Perplexity AI). Several general assistants also offer a search-first mode. |
| Suite-embedded assistants | Copilot in Microsoft 365 (Microsoft) · Gemini in Google Workspace (Google) |
| Real-time and social tools | Grok on X (xAI) |
| Task agents | Agent modes offered within several of the general assistants above |
| Open-weight models for self-hosting | Gemma (Google) · Llama (Meta) · Mistral models (Mistral AI) |
Listed alphabetically within each family. Names and makers as of Oct 2026. Products, names, and features change often.
Numbers match the bracketed citations in the text. Also online at xotavo.com/sources.