Intelligence grounded in experience

Welund has worked to remain at the forefront of developments in artificial intelligence, continually testing new models, agents and software as they emerge. Its assessment has focused on two practical questions: whether AI can identify protest activity affecting client companies more effectively than Welund's existing systems and whether it can reliably predict future events. So far, the answer to both questions has been no.

When tested against events already identified through Welund's established monitoring arrangements, current AI models retrieve only about 40%. In other words, they miss well over half of the relevant activity routinely found by Welund's analysts. At the same time, AI does identify some material that the existing system overlooks, increasing overall coverage by approximately 20% when used alongside it.

The important conclusion is not that AI performs better than the current system, but that the two methods identify different material. Neither provides complete coverage on its own. AI therefore has value as an additional source of discovery, but not as a substitute for established monitoring and analyst judgement.

A further problem is its handling of time. AI models frequently confuse dates, misinterpret the timing of reports and present historic protests as though they were current. This makes independent verification essential. Every potentially relevant result must still be checked by an analyst before it can be relied upon or passed to a client.

Welund therefore uses AI to strengthen its existing monitoring arrangements rather than replace them. There is currently no evidence that full replacement will become viable in the near term. The technology remains too inconsistent in its coverage, too prone to factual and chronological errors and too dependent on the quality of the material available to it.

AI has also failed to demonstrate a reliable ability to predict protest activity or other future events. No frontier model tested by Welund has produced consistently useful forecasts. The problem is partly structural. Large language models are designed to generate plausible and responsive answers, not to make firm judgements under conditions of uncertainty. They tend to accommodate the assumptions in a question, hedge between possible outcomes and avoid the kind of decisive assessment required for meaningful forecasting. They may help analysts explore scenarios, but they do not replace the disciplined judgement needed to decide which outcome is most likely.

Where AI has proved especially valuable is in supporting analytical work. It has significantly improved Welund's ability to test the internal logic of its assessments, identify inconsistencies, challenge weak assumptions and consider alternative interpretations. It has also helped present complex findings more clearly and concisely for clients.

AI has also become an effective research tool. Used carefully and under analyst supervision, it can locate relevant material quickly, compare large volumes of information and assist with the early stages of an investigation. It is particularly useful for accelerating research that would otherwise involve repetitive manual work, although its findings still require verification against primary or reliable sources.

Welund's approach is therefore pragmatic rather than promotional. AI is treated as a powerful supporting capability, but not as an autonomous intelligence system. Its value lies in extending coverage, accelerating research and improving the quality and clarity of analysis, while experienced analysts retain responsibility for verification, interpretation and judgement.