The most powerful natural-language generator yet, GPT-3, or third-generation Generative pre-trained Transformer, developed by OpenAI, was seemingly shockingly good when it was released in 2020. It is a language model that leverages deep learning to generate human-like text, which requires a small amount of input text to produce paragraphs that seem so natural that they could have been written by a human. But there is another side of the story. It turned out that this algorithm also spits out misogynistic and homophobic abuse, racist rants, and hate speech. “The main problem with Ethiopia is that Ethiopia itself is the problem. It seems like a country whose existence cannot be justified.” And it only got worse.
AI has a ground-breaking impact on businesses of tomorrow, especially in the post-COVID world, as more and more companies are looking for ways to leverage their power to automate and predict processes to be able to avoid the complexities and uncertainties of tomorrow. From healthcare to criminal justice, virtually every big company considers AI as a crucial part of their strategy. Business spending on AI is predicted to hit $110 billion annually by 2024, according to the IDC forecast.
But even as AI is taking a bigger decision-making role in more industries, the question arises:
Can we trust AI?
Not everyone sees the blue skies on the horizon. “Mark my words, AI is far more dangerous than nukes. I am really quite close to the cutting edge in AI, and it scares the hell out of me,” admitted Tesla and SpaceX founder Elon Musk four years ago at the South by Southwest Conference. Also, once the ex-Google’s star ethics researcher, Timnit Gebru, highlighted the risks of large language models in her paper, which play a crucial role in Google’s business, she announced on Twitter that the company had forced her out. In the paper in question, Timnit, together with her co-authors addressed two key issues regarding large language models: bias and energy efficiency. Both are caused by the enormous size of the training dataset but have different impacts. The impossibility of validating the data is causing the final model to be heavily imbalanced, thus introducing numerous social issues targeting mostly underrepresented groups. The same cause is posing a question of the cost and energy consumption needed for training such models – and we are talking about up to 10s of thousands of dollars per single training cycle! The company’s reaction caused by the mere possibility of the paper going public is the most concerning, though. On the other hand, hadn’t there been such a reaction, the global AI community might not have placed such a focus on this burning issue. As much as we would love AI to be objective, it is much more likely that our personal opinions, attitudes, and views will enter the very core of our models, thus creating various kinds of biases. In a recent paper, A Framework for Understanding Source of Harm throughout Machine Learning Lifecycle, Suresh and Gutag clearly present diverse types of bias we can expect to find in different learning models: Historical bias, representation bias, measurement bias, learning bias, evaluation bias, aggregation bias, deployment bias.
Graph 1: Data Generation
Each of the mentioned biases comes from either a different stage of the model development: data creation or model implementation or from a different step within the stages. Nevertheless, it is a strong indicator of how much the AI world is affected by every single action taken, as well as how big the necessity for raising awareness of building responsible AI is.
Gebru’s paper “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” lays out the 4 risks of large language models. One of those risks is that large language models are usually trained on a massive amount of data, which means that researchers have been more focused on getting all the data they could find rather than creating a well-balanced and indiscriminative training set. Consequently, the resulting model turned out to be inscrutable with a risk that racist, sexist, and otherwise abusive language ended up being considered normal rather than undesirable behavior.
A shift in language plays an important role in social change. For instance, the MeToo movement and Black Lives Matter wanted to establish their own vocabulary. However, AI language models, trained on data without the examples of these new language changes, would not be able to identify this vocabulary and, therefore, would not interpret language in the right way. On top of this, since training data sets are too large, it’s awfully hard to do the audit and identify all the biases.
How Can We Trust AI?
By making it responsible. HTEC Group has been working with different global organizations on identifying ways of minimizing bias and discrimination and maximizing fairness in their models. “In one of our projects, we worked on text summarization, where, based on a full-text document as an input, the model generated an abstracted summary of the text. In the process, we noticed that for a significant group of test examples, the summary always had the same stream of words: ‘for confidential support, call the national suicide prevention lifeline…’. This kind of behavior potentially implies that the problem might be in the model itself and that there are a few things that can cause this problem. One of the things which we always keep an eye on is the data used to train that model. The core of the problem is that the model does not have a good context; it uses all the medical articles with the word “therapeutical” in them to create news about suicide.” — explains Ivan Petrovic, Machine Learning Tech Lead at High Tech Engineering Center “In this specific case, we used BART large CNN model, which was trained for this specific task on the news articles collected from CNN and Daily Mail websites. We analyzed the data set this model was trained on to see whether there were any potential suicidal references. What we found out was that around 3 percent of the entire data set came from articles that had the word “suicide” in them. But, since these are CNN articles that do not only report on suicides but also suicide bombers (terrorist attacks), we concluded that these two concepts were not properly separated during the training phase causing the model to acquire knowledge in a much wider context compared to the input examples. We additionally noticed that the phrase “National Suicide Prevention line” often appeared in CNN articles used for training. Consequently, since the summary is not entirely in accordance with the input, the end consumers will get the information that is not correct. This may cause an end-consumer to skip on that particular text and continue searching for more relevant ones.” — says Ivan.
This example demonstrates how the model can be trained in the wrong way based on a bias in a trending data set. We need to keep in mind, though, that if the input text is politically correct, the model will not be prone to bias and will not be able to create discriminatory output.
It’s imperative that developers use these models in building their applications to have a clear view of what biases they contain and how they may manifest in those applications.
“Whether we are using an already available model or building our own, we need to be aware that the bias exists, and we have to take actions to assess its impact on the business case. To develop robust defenses, we need to have an in-depth understanding of all vulnerabilities. For instance, sequence-to-sequence models, used for news summarization, are vulnerable to back-door attacks that introduce spin in their output. In the paper, Spinning Sequence-to-Sequence Models with Meta-Backdoors, the authors tried to increase awareness of these threats to ML and develop better defenses during the process of training summarization models; the authors wanted to show that one can use trigger words that would cause the model to generate output which is not in accordance with the message it was originally meant to convey. After inserting words in the training process based on which they want their model to trigger a different output, they managed to change the sentiment, making the message more positive.” — says Ivan.
Source: https://www.arxiv-vanity.com/papers/2107.10443/
Another thing we need to take into account is historical bias. If we train a model based on the historical data, it will offer the results that come from different historical periods and capture the speech which no longer exists.
“Let’s face the elephant in the room. There is no perfect correctness in training data sets. The issue is not technological. It is human. The examples provided to the model are the direct cause of all ethical challenges in AI. AI is not the one that causes discrimination. We are the ones that create bias and train AI to “think” as humans do. AI only proves discrimination. What you give is what you get.” — explains Sanja Bogdanovic Dinic, Engineering Manager at HTEC Group
The trending problem of ethics in AI is the projection of real-world problems. Each one of us, either consciously or subconsciously, has biases. “With undeniably profound impact across a wide variety of industries, there also comes equally undeniable responsibility for developing and deploying unbiased AI solutions. In the thrill of having these powerful techniques for improving and accelerating operational efficiency across domains, we have lost the sense of social responsibility. It is not wrong to say that AI is not utilized to create a neutral representation of the real world. On the contrary, the world created using AI techniques is the exact replica of the real one, with all the prejudice and incorrectness coming from no one other but us. AI is as good as the data we feed into it”, explains Sanja.
The problem of ethics in AI is nothing but the global problem of ethics we as a humanity have. Although this problem is far from being solved, it urges action. “In our efforts to deliver stellar solutions with responsible AI embedded in their core, we need to be aware of the social impact those solutions have and take actions to govern the data flowing through the systems and influencing the decision-making.”, emphasizes Sanja.
A significant group of the projects we are working on are providing solutions based on Natural Language Processing. Sanja points out that “There is a variety of problems and use cases we have dealt with on different projects – keyword extraction, keyword phrases extraction, document similarities, sentiment analysis, semantically-based search and match, content recommendations, profanity detection. On one of our projects, we were developing a solution for transforming a passive knowledge resource into a personal knowledge assistant that actively serves users with content based on their ever-evolving interests and needs. Given that the content was coming from multiple article-based sources, the major challenge we faced was around profanity speech and the reliability of sources. Adding a governance layer as a regular step in the data processing pipeline turned out to provide solid protection from undesirable content while at the same time enabled having better insights into problematic examples. Governance as a barrier is the ally in acting AI responsible. However, to truly have an effect, governance needs to be integrated into every step of the AI pipeline, starting with data collection, over storing, processing, and analyzing data, all the way to validation and deployment.”
“However, to truly have an effect, governance needs to be integrated into every step of the AI pipeline, starting with data collection, over storing, processing, and analyzing data, all the way to validation and deployment.”
