algorithm fairness label

what is it about?

The “algorithm fairness label” is one step towards trustworthy algorithms. Like a seal of quality it ranks the fairness of algorithms according to communication.

what is to be achieved with it?

It is intended as a pressurising tool to ensure that companies develop their algorithms more fairly and transparently.

why is it important?​

It shows a strategy where not the companies but an objective institution can decide if ethical rules are appropriate.

who is it for?

It is intended to offer the user of service platforms a 
comparable evaluation of different portals. Clear communication has to be provided so that the user is able to decide which service he wants to use.

which algorithms get scored?

It focuses on companies offering direct services to its customers. For example credit scoring services, using algorithm processing to rate creditworthiness or predictive policing systems, estimating the potential threat in urban areas or posed by individuals.

how does it work?

The seal evaluates whether bias is present and to what extent it is communicated. Therefore it is divided into two parts: Fairness and communication.


To concentrate on communication, it first has to be figured out if bias appears in the system. Two procedures are used for this purpose.

1. real outcome data are examined with reverse engineering. A method to analyze the results and to redevelop the algorithm.

2. additional data is created by test runs. By changing individual parameters, it can be determined whether the outcome data also changes. For example does gender, age, race or place of residence influence the results.


After analysing the system’s bias we test how information is communicated. Depending on the context different rating criteria are required. We developed guidelines to detect, analyse and score the communication. The communication part has a stronger weighting in the label. Don’t forget: you can’t prevent bias completely. The question is how it is communicated.

how can communication
be evaluated?

In order to evaluate communication in terms of fairness, we have developed five communication guidelines. To make the guidelines more understandable we have added sample questions in the graphic below.

1. Sensitivity

To make an algorithm fair, it must be clearly communicated that you are dealing with automated decision making processes.

  • Are the goals, purposes and intentions of the algorithm clear?

  • Did the system communicate that a decision, content or advice is the result of an algorithmic decision?
  • Does the algorithm clearly communicate that its social interaction is simulated and that it has no capacities of “understanding” and “feeling”?

2. Transparency

Informations have to be retraceable and communicated transparently.

  • Does the algorithm give  alternatives to the shown results?

  • Does the user get an explanation why the algorithm made a certain decision?

  • Does the operating system give  informations about its certainty?

3. Accessibility

Information has to be fully accessible. Unpleasant information has to be communicated equally.

  • Is information shown in a logical hierarchy?

  • Is certain information hidden, to not be found easily e.g. appears as very small text?

  • Are negative aspects of the decision making process hidden in the interface?

4. human judgement

Humans have to be integrated into the decision making process of algorithms.

  • Does the user have the possibility to negate certain assumptions?

  • Is there a possibility to interact with the algorithm to check certain decisions?

  • Does the user have an opportunity to report inaccurate decisions made by the algorithm?

5. ethical importance

Potential consequences caused by the algorithm must be assessed.

  • Does the algorithm affect human autonomy by interfering with its decision-making process in an unintended way?
  • Could the algorithm’s decision influence the users decision-making process?
  • What harm would be caused if the system makes inaccurate predictions?

what does
that look like?

To better understand our evaluation process, here is an example of how it could look like. Imagine the evaluation of a credit scoring system. In the first step, the fairness of the algorithm is examined to determine whether bias is present. Let’s say the analysis shows the following bias in the system: 

Women between 40-50 earning less than 80,000 US $ per year receive a worse credit offer than men with same conditions.

The place of residence influences the credit offer regardless of the annual income

There are indications that lower typing speed is reducing the credit offer in general.


The system is obviously biased and therefore gets a bad rating in the fairness area.


But what about the communication. Can it even compensate for possible errors? How does the system communicate that the credit is denied? For this purpose, we apply our communication guidelines with the following sample questions:

1. Sensitivity

Is it clearly communicated that the decision is based on automated decision making processes?

2. Transparency

Are the reasons for the decisions obvious?

3. Accessibility

Is there an indication on the certainty on the result shown?

4. human Judgement

Is there a possibility to report inaccurate decisions?

Is it possible to contact an employee, to jointly question the result?

5. Ethical Importance

Which consequences could be caused by not giving the credit offer?

positive example

You may not like that you do not get the credit, but you can understand why it came to this decision.

The positive example fulfils many points of the communication guidelines, therefore the seal shows a good rating in the communication section. The overall rating of the seal can therefore be positive even if the Fairness section was poorly rated by algorithmic bias.

negative example

Why did it come to this result? You can only accept it, but do nothing about it.


Bad fairness.
Bad communication.
Bad overall rating.
Seems obvious.

The positive example fulfils many points of the communication guidelines, therefore the seal shows a good rating in the communication section. The overall rating of the seal can therefore be positive even if the Fairness section was poorly rated by algorithmic bias.

Bad fairness.
Bad communication.
Bad overall rating.
Seems obvious.