Christian Behr: Good AI Models Do Not Fix Bad Prompting

Good AI models do not fix bad prompting

Since the public release of ChatGPT on 30 November 2022, professionals have developed a strange habit. They talk about models nonstop: New releases, larger context windows, reasoning, agents, Copilots, benchmarks and model comparisons. Which model is better for drafting, what works best for claims, which model “understands” technology and which model is more “creative”.

The Wrong AI Debate

Most of this discussion misses the operational problem. The problem is not that professionals do not have access to AI. The problem is that many of them still do not know how to instruct it in a professional workflow. And in patent work, this really matters.

Consumer prompting is forgiving. A user asks for a travel plan, a recipe, a polite email, or a summary of a newspaper article. If the answer is slightly generic, nobody dies. Nobody files an overbroad claim with no technical support. Nobody creates a disclosure record that later collapses under scrutiny. But our Patent work is different.

Why Patent Work Is Different

Let us use “Invention harvesting” as an example, it is not conversation entertainment. It is structured technical extraction under legal and commercial pressure. The task is not to receive a pleasant answer. The task is to preserve the invention, identify its technical contribution, separate facts from assumptions, and avoid contaminating the record with unsupported language.

That is not a wording problem, it is a process problem.

The Illusion of Professional Output

A better AI model can make a bad invention record look more professional, and that is precisely the danger. Many users still believe that prompting means clever phrasing. They ask for “the best prompt”. They want a secret formula. They add more adjectives. They write longer instructions. They tell the model to “act as a senior patent attorney”. Then they paste a messy transcript into the system and ask it to summarize the invention and draft claims.

This is not professional prompting. This is wishful outsourcing.

Why Extraction and Drafting Must Be Separated

Consider a typical invention disclosure interview. The inventor explains a new cooling arrangement for a power electronics module. There is talk about a channel geometry, a pressure drop, a manufacturing constraint, and a temperature improvement. The inventor assumes that everyone understands which dimensions matter. The patent engineer assumes that the prior art problem is clear. The AI receives the transcript and generates a well-written invention disclosure.

It sounds good. But the channel dimensions are missing. The flow regime is not defined. The specific difference over the nearest prior art is buried in a side remark. The model quietly fills the gaps with plausible engineering language.

Now the output looks professional. That does not mean it is reliable.

In fact, it may be worse than an obviously incomplete disclosure, because the polished language hides the missing technical substance. The human reviewer relaxes. The document has headings. It has a technical field. It has advantages. It has embodiments. It has the smell of patent drafting.

But the invention has not been harvested. It has merely been dressed.

Prompting Is Process Design

A strong model does not turn chaotic input into disciplined IP work. It only processes the chaos faster. This is why extraction and drafting must be separated.

Extraction is the controlled collection of technical facts. What is the feature? Where is it shown? Who said it? Is it in the drawing? Is it in the prototype? Is it measured? Is it assumed? Is it essential? Is it optional? What is the technical effect? Against which prior-art arrangement does that effect matter?

Drafting is something very different. Drafting is the legal and linguistic construction of a patent document based on verified material.

Mixing these two steps is one of the most common mistakes in AI-assisted invention harvesting. The prompt “summarize the transcript and draft claims” may feel efficient. It is usually premature.

The correct order is less glamorous:

•  Extract
• Verify
• Structure
• Challenge
• Then draft.

Prompting is not the art of asking nicely. It is the discipline of preventing the model from doing the next step too early.

A structured workflow forces the model to extract features into a controlled table or section. It requires source references back to the interview, drawings, lab notes, or inventor statements. It separates confirmed facts from assumptions. It flags missing parameters. It asks follow-up questions before drafting begins.

That is professional prompting. It is not longer prose. It is better sequencing.

The Prior-Art Challenge

Another example is prior-art differentiation. An inventor says, “Unlike existing systems, ours reacts faster.” A weak prompt lets the AI convert that into a broad statement about improved responsiveness. A better workflow asks: faster than what, measured how, caused by which feature, under which operating condition, and compared with which known system?

Without those questions, the AI may produce a claim that includes the general control architecture but misses the actual distinguishing feature. The patent application then becomes broader in language and weaker in substance.

That is not an AI failure. That is a workflow failure.

The model cannot preserve a distinction that the process never forced anyone to define.

Hidden Assumptions and Missing Facts

The same applies to hidden assumptions. A model may describe a sensor arrangement as “configured to detect deformation”, even though the inventor only mentioned vibration. It may turn a manufacturing preference into a structural limitation. It may treat an optional embodiment as the core invention. It may smooth over uncertainty because natural language models are good at smoothness.

Source Discipline Matters: Smoothness is Not Evidence

In patent work, every important sentence should have a source discipline behind it. Either the statement comes from the inventor, the drawing, the measurement, the product specification, the prior-art analysis, or a deliberate professional conclusion. If nobody knows where the statement came from, it should not quietly migrate into the disclosure.

Professional AI use in IP therefore needs workflow logic. It needs staged prompting. It needs review gates. It needs source labels. It needs controlled extraction. It needs a habit of asking the model to identify what is missing, not only to write what sounds complete.

Better models increase fluency. Better prompting increases reliability.

This is the uncomfortable point in many current AI discussions. Professionals prefer to talk about model capability because that discussion is external. It puts the burden on vendors. It allows the user to wait for the next release.

Prompting discipline puts the burden back where it belongs: on the patent professional.

The Real Bottleneck

The real bottleneck in AI-assisted invention harvesting is not whether the model can produce a fluent paragraph. It can. The bottleneck is whether the professional process forces the correct information to appear before the paragraph is written.

A bad prompt asks the model to impress.A good workflow asks the model to expose the gaps.

That is the difference between AI-assisted drafting and AI-assisted self-deception.

The Limits of Better Models

Good AI models are useful. They are fast. They are patient. They can reorganize messy material. They can help produce clear invention disclosures and better first drafts. But they do not remove the need for disciplined harvesting. They make the consequences of undisciplined harvesting less visible.

That is why the next productivity gain in patent work will not come from another model comparison chart. It will come from professionals who finally understand that prompting is not typing instructions into a chat window.

It is the architecture of the work itself.

Conclusion

The future of AI in patent practice will not be decided by who has access to the newest model. It will be decided by who builds the most disciplined process around it. Models will continue to improve. They will become faster, more knowledgeable, and more fluent. Yet none of these advances can compensate for missing facts, undefined distinctions, unsupported assumptions, or poorly structured invention harvesting.

In the end, the quality of the patent record still depends on the quality of the workflow that produces it. Professionals who treat prompting as process design rather than clever wording will extract better inventions, ask better questions, and create more reliable patent applications. The competitive advantage is not the model itself. It is the discipline with which the model is used.

IP Lawyer Tools by Martin Schweiger