Update 10-Sep 25: Added links to papers about Thompson’s work on Nigerian driving apps.

I have two PhD students who are working on apps that use NLG and AI to encourage safer driving: Jawwad Baig has developed an app for UK drivers, and Iniakpokeikiye Thompson has developed an app for Nigerian drivers. Both Jawwad and Thompson are building on work that Daniel Braun did in the SaferDriver app (DOI link) he developed at Aberdeen.

Jawwad and Thompson have recently carried out experimental evaluations of their systems, and in both cases they are seeing a statistically significant decrease in unsafe driving incidents! Which is very exciting, I think this technology could have a real impact on a major cause of premature death.

Concept

Safe driving apps are a type of behaviour change. The basic concept is as follows:

• Use GPS and other sensors to monitor driving behaviour (speed, acceleration, braking) and driver behaviour (eg, mobile phone usage).
• Analyse this data to identify unsafe driving incidents such as speeding (using resources such as OpenStreetMap and Google Maps for speed limit data).
• Extract key messages from the data.
• Communicate these messages to the user in a psychologically appropriate way to encourage behaviour change.

Of course there are many driving apps on the market (eg, Damoov), and also many insurance companies provide apps to their clients. Most of these rely on dashboards and visual feedback; many provide financial incentives (eg, reduced insurance premiums) for safer driving. Most of these apps have not been carefully evaluated, but there are some good evaluation studies in this space, such as Ebert et al 2025.

Our goal is to see whether textual feedback can help, by using AI and NLG to provide deeper insights about driving behaviour and problems, as well as useful background information (eg, about laws and penalties). Since reading text is cognitively demanding and we do not want to distract drivers, we provide feedback after a trip (or on a daily/weekly basis), not in real-time during a trip. We do not provide financial incentives for safer driving.

The apps are evaluated by recruiting a set of drivers, and then

• Monitoring them for a few weeks (without using the app) in order to measure the baseline frequency of unsafe driving incidents for each driver.
• Asking the driver to use the app for a period, and monitoring frequency of unsafe driving incidents in this period.
• Collecting qualitative feedback from the drivers at the end of the experimental period.

As mentioned above, results look good, with both statistically significant reductions in unsafe driving incidents, and positive qualitative feedback highlighting specific cases where drivers changed their behaviour.

Safer driving app for UK

Jawwad Baig has developed a driving app for UK drivers. A preliminary version of the app was described in Baig et al 2022 (ACL Anthology). The app gathers data on speeding, braking, acceleration, and phone use. It also looks for speeding near sensitive areas such as schools, and highlights these as a special category.

Jawwad’s final system (no paper yet) improves on the prototype in many ways, of which the most important is user adaptation. A combination of rules and LLMs are used to adapt the text based on a number of user attributes, including age, driving environment, and the user’s preferred feedback style.

The system was evaluated as described above, with 30 drivers (varied age, gender, driving environment). Results were very promising, with unsafe driving incidents per mile falling by almost 50%. Qualitative feedback was also very encouraging, for example “The app telling me I was speeding near a school really hit home. I never thought about how many kids could be around at 3pm.”

Safer driving app for Nigeria

Iniakpokeikiye Thompson has developed a driving app for Nigerian drivers. The driving environment is very different in Nigeria, for example many drivers have never passed a driving test or had any education on safe driving. Also drunk driving (even by bus drivers) is common and often culturally accepted; it is illegal but enforcement is limited (paper). Data is also challenging; for example there were no data sets until Thompson created one (paper), and speed limit data was very difficult to get from Google Maps (etc).

When Thompson collected his data, he asked drivers to self-report whether they had drunk alcohol. Since this is culturally acceptable in Nigeria, the expectation was that the self reports would be accurate (this would not be true in UK), and Thompson used this data to train a classifier to predict from raw driving data whether the driver was driving under the influence of alcohol (paper).

Thompson’s app (paper) uses the alcohol classifier to add comments (where relevant) about the dangers of driving under the influence. Since many Nigerian drivers have limited knowledge, it also includes educational tips. The evaluation has just finished and he is still analysing data, but there is a statistically significant reduction in unsafe driving (although not as dramatic as Jawwad saw) and also qualitative comments highlighting specific cases where the feedback helped drivers.

Future

I have been involved in a number of behaviour change projects over the years, and driving feedback seems to be the most successful. Partially because the data is relatively straightforward to get (in contrast to dietary behaviour change, where getting good data about what people eat is difficult), and partially because addiction is usually not an issue (in contrast with smoking). The potential to help people is also enormous; driving fatalities are one of the leading causes of deaths for adults, especially young adults (WHO).

There are of course caveats and limitations. In particular, we know that a lot of health interventions lose effectiveness over time, and Jawwad and Thompson only measured impact over a period of weeks, they did not assess whether impact persisted over a period of years; it would be useful to investigate this in future work.

Both Jawwad and Thompson are talking about continuing to work on their apps after they finish their PhDs, and commercialising or at least deploying the apps, so that their potential to help people is realised. Thompson is also talking to people in other African countries about his work.

So I am excited! The apps are not bleeding-edge AI or NLG, but they do show how NLG and AI can help tackle a major worldwide societal problem, which is also one of the UN’s SDG goals.