Taxi apps collect thousands of data every day. Both drivers and customers share their location and movement data via the app. The system knows where a driver is. But also what the distance from the last customer to a potential new customer is. Everything is used to match supply and demand as quickly as possible.

The taxi app uses an algorithm where drivers move from one ride to another as quickly as possible. The expected time for the current trip is calculated and the expected arrival for the next one is already passed on. Here, choices are made by the algorithm when selecting the rides. It optimizes for to have enough time to make it to the next one, but not so much that the driver is idle. The assumption here is that both customer and driver want to wait as short as possible.

The algorithm is set up to optimize for scheduling as many rides as possible. This way, drivers don’t have to spend time finding new rides themselves. In theory they can earn as much money as possible and work continuously and for a long time. In practice, drivers can experience a great deal of pressure, because they are constantly on call. The algorithm could also be set up so that drivers take more breaks. This would give them better working conditions.

The self-learning algorithm was trained using data from 30,000 sample files. Some of these were files handled by officials that had been confirmed to be ‘right’ or ‘wrong’. Another part of the files came from a ‘blacklist’ of potential fraudsters. This introduced a non-representative risk estimate into the model; it was trained with files of which it was partly unknown whether they were ‘right’ or ‘wrong’. This creates a selection bias. When would the outcome of the risk assessment have been more reliable? If the algorithm were trained on representative sample files, of which it was checked whether they were indeed ‘right’ or ‘wrong’.

The algorithm used criteria based on a list of indicators chosen by employees of the department. Unfortunately, that way human biases can be incorporated into the algorithm. For example, the characteristics ‘Dutch/non-Dutch’ and ‘income’ were seen as indicators for fraud. To counteract this kind of bias, criteria for selection must be verifiable and public. This does not automatically make them perfect or objective, but it provides the opportunity for adjustment and improvement.

In the Dutch ‘benefits affair’ there was no mechanism to check whether the risk assessment for fraud was actually correct. As a result, thousands of people were duped for years, without the system being adapted. When using self-learning algorithms, the outcome should always be tested randomly. So that the method can be corrected when necessary.

Everything you post, share or write on social media is stored as data. Recently, researchers from Oxford and the Alan Turing Institute created a database of 4,000 examples. The database is available to everyone worldwide: Hatecheck. This dataset was developed through a combination of 29 different tests and manually determining whether something is classified as hate speech or not.

The researchers interviewed employees of 16 different non-profit organizations working on hate speech online. Based on these interviews, 18 different types of written English hate speech were defined, including derogatory and threatening language. In addition, 11 scenarios were defined that can mislead AI. For instance vulgar words in innocent statements or the use of offensive language in protest messages, so-called ‘counter-speech’.

The criteria were tested on three commercial services that use content moderation and collect responses from their readership. One of the conclusions was that hate speech detection by AI still has many challenges. If you moderate too little, you will not solve the problem of online discrimination and intimidation. Moderating too much could censor the kind of language marginalized groups use to empower and defend themselves.

Data for climate models are a combination of physical ‘certainties’ and variables that depend on ‘uncertainties’. Certainties are, for example, how molecules react to temperatures (water always freezes at 0℃). An uncertainty is, for example, how much CO2 will be emitted in 50 years. These data are combined in a predictive model. This creates a picture of the effect of the ‘uncertainties’ on the climate and on sea level rise.

What will happen tomorrow is unknown, but that the sun will rise is certain. Certain laws of physics will still apply in 100 years. A predictive model combines ‘certainties’ with factors that are known to be uncertain. The more factors, the greater the number of possible outcomes. By having models calculate all these different scenarios, researchers gain more insight into the most likely scenarios. In this way, the predictions become more and more reliable.

Predictive models show a range of possible outcomes for the climate. For example, there is great certainty that the sea level will rise, but to which extent exactly is more uncertain. After all, this partly depends on factors that will come into play in the future and which are currently unknown. In the Netherlands we now take into account a sea level rise of 30-80cm in the year 2100 in the most optimistic scenario. And up to 1.2m in the least favorable case. In both cases, the rise greatly increases the risk of flooding.

Weather data and climate data have many similarities. For example, both use temperature, humidity, ocean and air currents. The big difference is the time scale. Climate models use data that show what the weather was like centuries ago, such as tree rings and carbon dating. Weather models use data from short term events. For example, they focus on clouds and temperature differences between layers of air. For climate, these individual events are not relevant. Climate concerns the average effect of cloud cover in the longer term.

Both weather models and climate models are complex. They take into account many different influences. The daily course of the sun has a major influence on the weather. A winter day may have fewer hours of sunshine than a summer day. So is it colder? Possibly, but not always. A weather model must be as accurate as possible for a very specific place and time. For the climate, local, daily or seasonal changes count less. Climate models calculate on a different scale and thus do not indicate what the weather will look like in a particular place in 100 years. They look at the effect of this weather on global conditions, such as the range of temperature and sea level.

A weather model contains parameters that consist of certainties and uncertainties. For example, air currents and wind directions can be predicted fairly well, but the formation of clouds is much more unpredictable. Therefore, tomorrow’s weather forecast in your city is not always accurate. However, this weather has very little influence on the climate in 100 years. For this, models use more stable factors and deal with other types of uncertainties. For example, tomorrow’s weather may be more difficult to predict than a global temperature rise of 1.5℃.

Climate models use different types of data, for example about the composition of the atmosphere and temperature of ocean and land. Also, a lot of data is collected after volcanic eruptions. Such as how much ash is released, how it spreads and which other substances end up in the air and on land.

The change in the atmosphere and its effects on the climate can be well predicted for volcanic eruptions using climate models. For example, based on the size of the eruption and the composition of the ash. After all, the models already take these kinds of factors into account. However, the models do not have the data to predict when and where a volcanic eruption will take place.

Climate models can mimic the consequences of a volcanic eruption, but cannot predict when they will occur in the future. This would create a great deal of uncertainty in the model. Therefore, Asha chooses not to include them in predictive models. She involves these kinds of defined events when they occur or have happened. In this way, a scenario is updated based on current events, and the model itself becomes more powerful.

The algorithm uses data to make choices in the checklist. In this case, there are two types of data; the patient’s data (such as weight, age, underlying conditions) and the data used to create the algorithm (such as scientific research, available medication, hospital protocols). For both types, choices have been made about which data is and is not included. For example, whether socio-economic status such as income or origin of the patient is relevant. But also about how far back the data goes into the patient’s medical history. Furthermore, the data from scientific studies must be up to date enough.

An algorithm is a series of calculations, where each next step depends on (adjustable) settings. These settings are made by humans and are based on assumptions and choices. For example, it must be decided together with doctors which liver function is appropriate for which dose of pain medication. By setting these steps as completely as possible, an algorithm is created that takes many different medical aspects into account. An algorithm with the right settings and relevant data supports the doctors and nurses in the ICU.

The Checklist app helps Joyce and her colleagues make decisions about ICU patients. The algorithm provides personalized information based on choices and settings made in advance. This reduces the chance of human errors or mistakes. It is important to gain insight into the data that has or has not been used, and the assumptions that have been made. Adjusting settings in response to new insights is crucial when using this app responsibly.

When training the algorithm, Jin used as many X-rays as possible from different departments of the hospital. An algorithm does not ‘see’ a photo as a whole as we do, but as a collection of pixels of different intensities. The data for the algorithm therefore does not consist of a stack of photos, but is instead a long series of numbers from which the algorithm must learn to extract the relevant information.

With ‘black box’ AI, an algorithm teaches itself which information in the series of numbers is relevant. The algorithm is trained with a variety of photos that are marked as ‘bone fracture’ or ‘no bone fracture’. This allows the algorithm to recognize patterns in the two categories. After the training phase comes the testing phase. With a series of new photos, the algorithm is asked to choose the category itself. It is unknown to Jin which pattern the algorithm learned to recognize. If Jin wants to know, he can choose so-called ‘explainable’ AI. The learning process is then divided into steps and the algorithm shows which patterns it recognizes at each step.

Jin’s ‘black box’ algorithm turns out to be very good at recognizing photos with bone fractures. However, Jin discovered through explainable AI that the recognized pattern has little to do with the medical data in the photo. Bone fractures are most common in the emergency room and make up a relatively large proportion of the photos. Moreover, these are often taken in a slightly different way than in the other departments. The algorithm recognizes this pattern; different background, position etc. So the algorithm needs to learn to skip the department-specific pattern. Only then is it really useful in identifying bone fractures.

Data from smartwatch apps can be collected in different ways.

• By observation: the data of all users who have given permission is stored. This data is therefore passively obtained and can be analyzed afterwards.
• By selection through a designed experiment. The researchers select the type of participants in advance based on their research question and actively approach them to participate.

In the first way, a lot of data can be collected with little effort. The second way takes more time and effort.

When a health insurer analyses data collected through observation, they must take into account that these data do not provide a complete reflection of society. After all, the data comes from a specific group. Namely; people who, for various reasons, use a smartwatch app. The active way of collecting data through selection takes more time and effort, but gives more chance of an objective answer to the research question. Because that way, also people who would not otherwise use a smartwatch app are investigated.

The health insurer concludes that people who use a smartwatch app are healthier and therefore may receive a discount on their health insurance. This may be true. However, based on observation data alone it is not possible to know whether the use of smartwatch apps makes people healthier. To establish a cause-effect relationship, data from a representative group of people is needed.

Smart energy meters collect large amounts of data, all day every day. The energy meter measures the level of electricity and gas consumption at any time of the day. In the app you can see exactly how much energy you have used and when; your usage data. Data about your type of home, your address, and personal data are also collected, to make sure that the data you see in the app is yours.

The app provides insight into your energy consumption. It only shows data without applying analysis to it. However, the digital collection of data does make it possible to conduct major analyzes and to recognize patterns. Analyzing the consumption data of all energy meters provides insight into how certain neighbourhoods, or perhaps even people, have arranged their lives. What time do you get up and what time do you go to sleep? Your energy consumption provides insight.

It is useful to look at your own consumption, but it is not self-evident for this data to be accessible to others. This type of privacy-sensitive data must be handled with care. For example through strict rules for collecting and storing data, so that personal data is separated from consumption data. When energy consumption data are combined and pooled they cannot be linked to your own home. In this way you ensure that the source of privacy-sensitive data is used responsibly.

TNO uses predictive models with cost optimisation. This makes it possible to assess the costs and technical characteristics of a wide variety of energy transition scenarios. The algorithms in the model use data about the costs and technical characteristics of many different technologies. Nuclear energy is one of the technologies included in the model.

The cost optimisation model is used to see how we can achieve a goal such as climate neutrality at the lowest possible cost. In all scenarios of this model, it appears that electricity from solar and wind plays a major role. A scenario supplemented with nuclear power appears to be less optimal. The job of the algorithm is to select for affordability. For that reason, the model does not opt ​​for, more expensive, options with nuclear energy.

Giving priority to cost efficiency, this model looks at how the Netherlands can best generate electricity in a climate-neutral future. Does this mean that nuclear energy is excluded from for (ideological) reasons? No, quite the opposite. It is up to politicians like Wim to decide what to do with the outcome of the TNO models. The financial aspect is only one of the considerations in the decision process of energy transition.

A model that calculates how to best create a natural gas-free neighborhood involves a lot of data. Data such as the type of homes, level of insulation, location for solar panels and possibilities for geothermal energy or heat networks, for example. Also relevant is whether the homes are owner-occupied or rental properties. In addition, Brechtje must choose which data to use exactly, for example; year of construction or energy label. Information about the year of construction of a house is reliable, but does not say much about the energy label. Energy labels may provide meaningful information about a home, but they are often outdated.

Making choices and making assumptions are an important part of building a model to find the most efficient way to become natural gas-free. For example, choosing to look at the year of construction, means assuming that modern homes are better insulated than old ones. If you opt for feeding the model with data on energy labels, the assumption is that this information provides a sufficient picture of the current situation. It is important to carefully consider and record these assumptions.

Which assumption Brechtje chooses, affects the outcome of the model. When using energy labels, for example, she already knows that the data can be outdated. The current situation has improved in many cases. How much insulation is still needed can therefore be overestimated. As a result, natural gas alternatives may seem less feasible than they are. The same problem occurs, but to a greater extent, when choosing the year of construction.