Frequently asked questions


A - Functionality and purpose

1. In two sentences: How does the PEAK-A work?

The PEAK-A is an AI-supported potential analysis for peak performance that consists of a test battery and a language effectiveness analysis. The combination of individual dimensions and psycholinguistics ensures the plausibility of the results.

2. What are the components of the PEAK-A?

The PEAK-A is a combination of personality test, motive profiles and leadership understanding with an AI-driven language effectiveness analysis.

3. What are the benefits of the PEAK-A?

The PEAK-A uncovers individual "blind spots", identifies potential learning areas and enables genuine self-reflection on the path to peak performance.

4. What distinguishes the PEAK-A from other personality tests and what are its special features?

Conventional established test procedures are based on an understanding of leadership from the 1950s and 1960s and no longer correspond in all aspects to the current challenges faced by managers. They measure so-called dispositional personality traits, which are not directly observable but manifest themselves in the behaviour of individuals. If you are aware of this, you will learn to deal with the behaviour, but you will not lose the personality trait.

The PEAK-A test theory also captures motives that can be seen as fundamental drivers for our behaviour. The scales have been transformed from elite sport and represent how people experience and perceive situations. These differences in experience not only influence our thinking, but also what a person can or cannot achieve in a particular situation.

Understandable and meaningful communication is a key to leadership success and co-operation. Therefore, language should be a key component in contemporary testing procedures. Understanding the impact of communication therefore provides important information for self-reflection and identifies individual learning areas. The impact of communication should not only be assessed on the basis of content. Linguistic impact is also characterised by formal language patterns. All of this is combined in the PEAK-A.

5. How should the results be read?

The results are always presented in such a way that the significance of the respective dimension for peak performance is explained in general terms and your personal results, including individual categorisation and characteristics, are presented. You can always recognise where you stand in comparison with the overall collective. The cup indicates where the representative peak performance panel is. You can derive your individual learning areas from this and have the opportunity for self-reflection.

B - Technology

6. How does the AI language impact analysis work?

PEAK-A speech analysis is an artificially intelligent technology that makes psychological deductions from spoken and written language. The technology was trained for this purpose in large-scale studies:

In a first study, a psychological model was developed that depicts language as holistically as possible with 29 mathematically tested constructs derived from the literature. In a second (ongoing) study, the technology was taught to recognise these constructs in language as independently of the situation as possible. To this end, millions of evaluations of texts were collected in various contexts and the small-scale analysis of the language was further refined so that the mathematical representation of the language could be brought together with the evaluations by the AI. A multi-stage modelling process ensures the quality of the statements.

7. What criteria does the algorithm use to search?

The analysis always takes place on the basis of a text. Spoken language is therefore automatically transcribed before being analysed. The PEAK-A speech analysis has learnt to "understand" speech by using a model with more than 110 million parameters. Example parameters are words, word combinations, grammar, the position and type (adjective, noun) of a word in the text and so on. The PEAK-A language analysis therefore does not search for previously defined human criteria, but is based on a mathematical model of language.

This holistic mathematical understanding of language enables machine learning to take place. During training, for example, PEAK-A Language Analysis is shown texts that should have a high or low value in one of the measurable results. These could be, for example, scientific articles for highly "intellectual" communication - and articles from tabloid newspapers for low "intellectual" communication. In addition, millions of evaluated text passages are shown to the technology. For these text passages, several raters have assessed whether they are "intellectual" using specially developed ltems. From all the information, the technology builds a model for "intellectual" and learns to recognise which aspects of the language (from the model with 110 million parameters) are relevant for the construct.

Once a model has been built, a data set (DEV data) is used to check whether the model works well. The DEV dataset consists of further ratings of text passages by several raters, which were assessed using the developed items. If the model is good or better than the previous one, the next step can be taken. If the model is not better than the previous one, it goes back into training. New training methods or other training data are then used. For example, the PEAK-A language analysis is shown high-quality newspapers instead of scientific articles, as it is possible that the technology has learnt too specific aspects of "intellectual" communication from the scientific articles. Alternatively, more ratings could be collected, etc.

Once the model has improved on the DEV dataset, the TEST dataset is used. The TEST dataset has a similar structure to the DEV dataset and also consists of millions of ratings for text passages. The decisive factor is that the DEV and TEST data sets are completely separate data sets. If the quality of the TEST dataset is as good as the DEV dataset, it can be assumed that the PEAK-A speech analysis has learnt a stable and good model. The model can be released. If the quality is better or worse than the DEV data set, troubleshooting begins. A significantly different quality compared to the DEV data set can have various reasons. Depending on the cause, further steps are triggered.

If a new text is added to the released model, the technology applies the information learned and checked from the modelling process described above and searches for the aspects from the 110 million parameters that have proven to be valuable for the statement about the result. The PEAK-A language analysis has thus learned what is relevant for a result by means of an elaborate modelling process, without being told from outside which parameters are relevant for a result.

8. What in the language contributes to the result, whether high or low value?

Words, word stems, word patterns, sentence structures and the combination of all relevant parameters are relevant for the analysis. PEAK-A language analysis does not look at linear relationships, but always includes the combination of all facets and the overall picture of the language when making statements about a characteristic. This means that it is not possible to pinpoint individual words that are the cause of a result and which alone need to be changed in order to influence the overall effect.

However, examples of contributing words and word combinations are always shown in the PEAK-A language analysis evaluations to make it easier to understand what was measured.

9. Who is behind the AI?

VIER GmbH is a company based in Hanover and an exclusive partner of Unfreeze People GmbH. At VIER, AI experts work hand-in-hand with psychologists and application experts. A large part of the work consists of research into and development of the technology. The other part of the work is advising on the optimal use of the technology with the aim of designing meaningful, individualised feedback, improving processes, recognising and editing patterns or enriching existing data sets.

10. What are the quality criteria? - How is the criterion validity of the tool proven?

Criterion validity indicates whether the result of a process is related to a relevant external criterion, for example better response rates to cover letters, higher analyst ratings in financial market communication, professional success or sales results.

The criterion validity of PEAK-A speech analytics was demonstrated in various customer projects with regard to performance-related key figures with the help of application-related analyses. For example, a conversation based on PEAK-A Language Analysis can improve the effectiveness and clarity of language and the targeted use of communication. At the same time, the results are linked to management evaluations and company success. More detailed information on this is available on request.

Criterion validation is constantly being expanded through ongoing evaluations in projects and simultaneous scientific studies.

11. Can I prepare for the interview?

The interview is designed so that you can simply get started. You do not need to prepare. The questions deal with general topics such as your typical Sunday (first question), hobbies, experiences, professional career and so on. It is not the content that is analysed, but your language structure. You can also skip questions if you don't want to answer them.

However, you should find a quiet place where you can talk on the phone without being disturbed. You should also allow enough time (approx. 15 minutes) to conduct the interview in a relaxed manner and without interruption.

12. How are biases prevented?

We pay attention to objectivity and representative distribution when generating the data.

In addition to the prediction model, we have built a second model that visualises a large proportion of the features (words, word combinations) that are included in the algorithm. This makes the prediction transparent and controllable.

Checking for systematic distortions in the results (e.g. by age or gender) is a central part of the quality criteria for our models. In this way, we ensure that no unnoticed biases are introduced as the model learns (better predictions). Both data generation and modelling are therefore subject to strict control mechanisms.

13. Can I manipulate the AI?

Your behaviour in different situations is intuitive. This means that you react to the situation and behave in the way you have learnt for this situation. Every person uses every communication style at some point in their everyday life. At the same time, you have a personal tendency to behave in different situations. For example, someone who tends to communicate in a very "authoritarian" way will always have a tendency to take the lead in conversations, lead the conversation and formulate appeals depending on the situation. Even if there are situations in which a reserved form of communication is deliberately chosen, the basic tendency remains. The longer the situation is, or the more spontaneously a question is answered, the more "natural" the reaction to this question will be.

Since it is also unclear to you which aspects of language contribute to a certain result and you therefore do not know which form of behaviour is "better" for a certain value, the reaction is usually very natural. Even if the parameters relevant to a particular outcome were known, it is almost impossible to stay on topic and at the same time focus on the small elements of language (sentence structure, word usage, ...) for a desired effect.

14. What does the system do better than a psychologically trained person?

Comparable input is crucial for analysing at the personality level. The technology is standardised to the PEAK-A speech analysis interview and ensures that all people are in the same standardised interview situation. This produces comparable data.

If the interview is conducted by a human interviewer, confounding factors such as sympathy/antipathy between interviewer and interviewee occur. The security of a fully standardised interview is not possible due to possible digressions controlled by the interviewer.

The advantage of the technology is the objective comparison with a representative sample in a comparable situation. This avoids effects such as the halo effect, in which individual characteristics of one person outshine others. Even trained psychologists cannot completely avoid these effects. The tool therefore offers an additional and completely different perspective that is significantly freer from classic biases. Of course, the EMOTION ANALYTICS interview should be supplemented by a procedure in which a personal impression of a candidate is gained (as with any other diagnostic procedure).

15. Where are the limits of AI?

The perceived limits of artificial intelligence are shifting very quickly. What we perceive as the limit today is already far beyond what would have been perceived as the limit 10 years ago.

Our actions are based on the assumption that artificial intelligence, in whatever form, can be made useful to humans. PEAK-A speech analysis can help to improve communication between people. At the moment, it seems inconceivable that technology can in any way replace the closeness and relationship between people in all its complexity. Rather, it can help people to better understand themselves and their relationships with others and, in the long term, become an integral part of everyday interactions by helping people to improve their communication with each other.

C - Data security

16. How is data security ensured?

Our partner VIER GmbH attaches great importance to the optimal protection of personal data. Therefore, there are extensive organisational and technical measures within the meaning of Art. 32 GDPR, which pursue the objectives of confidentiality, integrity and availability of data. They ensure that only data whose processing is necessary for the respective specific processing purpose is processed.