Citizens Are Writing Complaints with AI. Can Public Bodies Keep Up? Lessons from the UK's AI Complaint Wave and Tainan's AI 1999

People are now using AI to write complaints and requests, and the letters are getting longer and less accurate. Intake teams can absorb this with structured forms, AI summaries and human confirmation, rather than simply adding barriers.

The UK's wave of AI complaints: every letter is getting longer

The pressure on UK public bodies is not mainly a surge in case numbers. Each case is getting longer and harder to read.

The Guardian reported on November 8, as relayed by TVBS and South Korea's Seoul Economic Daily, that Deborah Evans, chief executive of Lawyers in Local Government, an organisation representing local authority lawyers, said the average complaint to a council ran to about two pages before chatbots became common, and the longest she knows of now runs to about 200 pages. She has also seen AI cite foreign precedents that do not apply locally, and even invent cases. A senior caseworker in a Labour MP's office estimated that about half of the letters the office receives show signs of being written by AI.

Data from Transport for London (TfL) offers another angle. According to a report to TfL's Audit and Assurance Committee, internal review requests on freedom of information decisions rose from 154 the previous year to 224 in 2025/26. The report notes an increase in requesters using AI tools to generate appeals, adding to the volume and complexity of managing them. TfL is considering a standard appeal template.

Not every increase can be attributed to AI, however. The Guardian reported that UK employment tribunals opened about 14,000 single-claimant cases from April to June, up 28% year on year, but Seoul Economic Daily noted in its account that it has not been established whether AI is driving the increase.

Researchers have a name for it: agentic flooding

The channels most likely to be overwhelmed are open, free online channels that accept unlimited free text.

Researchers from the Hertie School in Berlin and AI governance research organisations published a paper on arXiv in August, "Characterizing Agentic Flooding of Government Services", compiling 84 potential cases across 11 jurisdictions. According to UK legal publication Local Government Lawyer, the main findings are:

About nine in ten cases showed "qualitative" flooding, meaning longer and more complex requests; 60% showed "quantitative" flooding; about half showed both.
In 87% of cases, people used large language models to cheaply produce text that looks legally formatted, then submitted it by hand through existing channels. The researchers found no evidence yet of autonomous AI agents operating government websites directly.
The most exposed services are open digital channels that accept free text of any length, such as freedom of information portals, consultation platforms and public comment procedures.

The study also looked at how governments responded. In 17% of cases they added friction, such as fees, identity checks, per-claimant limits or IP blocking; in 25% they introduced their own AI tools, such as bulk analysis of public consultation responses. The researchers warn that fees and other friction are quick to deploy but disproportionately deter poorer and less digitally literate people. Lead author Chris Schmitz also said the method does not support causal or quantitative conclusions, and the true number of affected services is likely higher.

Taiwan's intake side: Tainan's AI 1999 builds human confirmation into the process

Local governments in Taiwan are already using AI at the intake stage, and designing the process so that a person confirms before anything is sent.

The Tainan City Government launched its "AI Municipal Service Engine" on November 5, with the AI 1999 smart customer service as its first validation case. According to China Times and Yam News, the system helps hotline staff convert citizens' calls into text, search municipal knowledge, summarise cases and suggest replies, and it can recognise place names spoken in Taiwanese. Hotline staff confirm the content before replying to citizens.

Two design choices stand out. The first is shared capability: the city says other bureaus can reuse the speech recognition, knowledge retrieval and large language model capabilities, reducing duplicate development. The second is resilience: the city, ITRI and AMD are building an architecture in which AI 1999 and existing platforms back each other up, and the city says it is enforcing data governance and security controls at the same time.

Taipei's 1999 hotline also introduced AI speech-to-text and an intelligent assistance system this year. According to Want Rich, Taipei's Research, Development and Evaluation Commission said the hotline handles about 100,000 calls a month, and the next phase plans to have AI route answers to frequently asked questions automatically to staff computers.

Why schools, agencies and businesses should pay attention

Any organisation with online applications, a feedback inbox or a customer service form is on the same front line.

University complaint and feedback inboxes, government petitions and information requests, and corporate complaint forms all share the same basic shape: open text input, followed by manual handling. When the cost of writing a long letter approaches zero but the cost of handling it does not fall, the pressure lands on the case officer.

Responses in the UK fall into three groups:

Adjusting the entry point: some ombudsman services have introduced word limits or online templates that cap attachments; TfL is considering a standard appeal template.
Using AI on the handling side: Wiltshire Council is exploring AI to summarise long emails and draft initial replies.
Holding the line: according to reports, UK government guidance tells civil servants to verify information provided by AI, not to input personal data, and to leave final decisions to people. Guidance from the Information Commissioner's Office (ICO) reminds organisations that compliant requests written by AI must still be handled under existing rules, and that they can ask requesters to clarify unclear requests.

How to redesign the intake process

The safer approach is to break free text into structured fields, let AI handle the remaining long text, and have a person confirm the result.

The UK Parliamentary and Health Service Ombudsman advises people who use AI to write complaints to state three things clearly: what happened, what impact it had, and how they want it resolved, and to check before sending that the content matches their own experience. Those three points can become form fields directly. With clear fields, applicants are less likely to paste in a long block of AI-generated text, and case officers see the key points first.

Long attachments and supplementary statements are well suited to AI summarisation and classification, but summaries should link back to the original passages so case officers can check them. When AI cites laws or precedents, the system can flag them as "to be verified" so that officers do not treat invented content as fact.

Finally, set data boundaries. Petitions and complaints often contain personal data, so summarisation and retrieval are best run in an environment the organisation controls, with access rights and logs consistent with existing systems. Tainan's approach of turning speech recognition and knowledge retrieval into shared services while emphasising data governance and security controls is a useful reference.

Recommendations for organisations

Inventory every entry point that accepts free text, including feedback inboxes, customer service forms and supplementary fields in online applications, and identify where long submissions are most likely to pile up.
Redesign forms with structured fields, at minimum separating "what happened", "what impact it had" and "how you want it resolved", and set reasonable limits on length or attachments.
Use AI for summarisation, classification and flagging legal citations, but always have a case officer confirm before replying, and make summaries traceable to the original text.
Draw personal data and security boundaries first, then decide whether AI runs in the cloud or on premises, and bring access logs into your existing security management.
Before adding barriers, assess the impact on vulnerable users. Fees or mandatory identity checks work quickly, but they can also shut out the people who most need help.

If your organisation is reviewing the forms and data flows of an existing online application or petition system, JIE HENG's government and campus information systems team can help take stock.