By Luis Lopez, AI transportation consultant, CEO of Go Hub.io Holdings Corp and subsidiaries, and host of the Freight Guru Podcast
An AI assistant can write a clear, well-organized answer that is completely wrong. It can cite a regulation section that does not exist, state a transit time nobody gave it, or fill in a missing container number with one that looks plausible. This behavior is commonly called hallucination, and anyone using AI in freight needs to understand it.
This article explains why it happens, where it causes real trouble in logistics work, and how to build habits and controls that catch it.
Why Language Models Make Things Up
A large language model generates text by predicting what should come next, one small piece at a time, based on patterns learned from a very large body of writing. It is not looking up facts in a database unless it has been connected to one. It is producing the response that best fits the patterns it has learned.
That design explains the problem. When the model has seen a fact stated consistently and often, the predicted answer is usually correct. When the question involves something obscure, recent, specific to your company, or simply absent from what it learned, the model still produces fluent text. A fluent answer and a true answer look identical on the page.
Several things make it worse:
- The model does not reliably know what it does not know. It may not signal uncertainty unless it is instructed to.
- Specific details are harder than general ideas. A model can explain what demurrage is far more reliably than it can state a particular terminal’s free time.
- Its knowledge has a cutoff. Rules, fees and procedures change, and a model without access to current sources will answer from older information.
- It aims to be helpful. Asked to complete a form with a missing field, a model may supply a reasonable-looking value instead of leaving a blank.
This is different from a predictive model that forecasts an arrival time from data, which can be inaccurate but does not invent text. The distinction is covered in our article on generative AI versus predictive AI in freight.
Where Hallucinations Hurt in Freight
Freight runs on precise identifiers, numbers and rules. A response that is almost right can be worse than no response, because it gets acted on.
Regulatory and compliance questions
Ask a general-purpose assistant about hours of service, hazardous materials requirements or customs rules, and it may give a solid summary. It may also invent a citation, blend rules from different jurisdictions, or describe a rule that has since changed. Compliance decisions should rest on the current official text or a qualified professional, not on a chat response.
Document extraction
AI tools that read bills of lading, rate confirmations and delivery orders save a great deal of typing. The risk is a field that is blurry or absent on the document being filled with a guess. A wrong container number, reference number or weight flows straight into dispatch and billing. See our overview of how AI is changing freight document processing for the wider picture.
Classification and customs data
Tariff classification depends on exact product details and legal notes. A model can suggest a code that sounds right and is not. Our article on AI in customs classification and entry prep explains why a licensed professional needs to own that decision.
Customer communication
An AI drafting a status update may state a delivery time or a reason for delay that appears nowhere in the shipment record. Once sent, that statement becomes a commitment in the customer’s eyes.
Rates and quotes
A language model asked for a rate without access to your pricing data will produce a number. That number has no basis.
How to Reduce the Problem at the Source
Hallucinations cannot be removed entirely, but the way a tool is built and used changes how often they occur.
- Ground the model in your own data. Tools that retrieve the relevant shipment record, contract or policy document and answer from that material are far more reliable than a model answering from memory.
- Require sources. Ask for the document and passage behind each claim. An answer that points to a specific page can be checked in seconds.
- Allow “not found.” Instruct the tool to leave a field blank or say it does not know when the information is absent. This instruction reduces invented values, though it does not eliminate them.
- Narrow the task. A model asked to extract six named fields from one document is easier to control than one asked an open question.
- Keep math out of the language model. Calculations such as charges, density or mileage should be done by regular software, with the model only explaining the result.
How to Catch What Gets Through
Controls at the source are not enough. The process around the tool has to assume some errors will arrive.
- Validate identifiers automatically. Container numbers follow a standard format that includes a check digit, and software can test it. Reference numbers can be matched against your system. Dates and weights can be tested for reasonable ranges.
- Cross-check against a second source. Compare extracted data to the booking, the order or the previous document in the chain. Mismatches go to a person.
- Use confidence thresholds. Where a tool reports confidence by field, route low-confidence items to review instead of accepting them.
- Verify every citation. If an answer references a regulation, tariff item or contract clause, open the source and read it. Treat an uncheckable citation as unverified.
- Put a person in front of high-stakes outputs. Quotes, customs filings, claims correspondence and anything with legal or financial weight should be approved by someone accountable.
- Sample the routine work. For high-volume, low-risk tasks, audit a random sample on a schedule. Rising error rates show up there first.
- Log the errors. Record what the tool got wrong and why. Patterns usually point to a fixable cause, such as a document layout it handles poorly.
Warning Signs in an Answer
Certain features should prompt a second look:
- Very specific numbers, dates or section references when you provided no source material
- An answer to a question about your own shipment that the tool had no data access to answer
- A different answer when the same question is asked again
- Details that are suspiciously tidy, such as round figures or a perfectly complete form from a poor-quality scan
- A confident tone on a topic where rules vary by carrier, terminal or jurisdiction
Setting the Right Expectations
As an AI transportation consultant, the first thing I look at when a team adopts a language tool is whether people understand that fluency is not accuracy. Staff who treat the output as a draft from a fast but unreliable assistant use it well. Staff who treat it as a reference source eventually get burned.
A short written policy helps. It should say which tasks AI may be used for, which outputs require verification, and who is responsible for the final result. The underlying principle is simple: the person who sends, files or approves something owns it, regardless of what drafted it.
For more practical guidance on using AI safely in freight and logistics, subscribe to the Freight Guru Podcast.
About the author: Luis Lopez is a Miami-based AI transportation consultant and logistics entrepreneur, the CEO of Go Hub.io Holdings Corp and subsidiaries, and host of the Freight Guru Podcast.
