Skip to Main Content

Artificial Intelligence

An image for artificial intelligenceAI Ethics

While artificial intelligence offers exciting possibilities, it's crucial to address ethical concerns surrounding bias, copyright, the potential for AI-generated hallucinations, and privacy violations.

Dive deeper into these issues by exploring the following tabs.

Bias

Recent research suggests that even after people stop using biased AI systems, those biases can linger and influence their own choices. This is especially troubling because biased data can amplify existing prejudices, leading to even greater inequality. In a study, participants who received biased recommendations from a simulated AI diagnosis tool were more likely to make biased decisions themselves later on. This points to the possibility of AI making discrimination and errors even more widespread. The research emphasizes the importance of proactively addressing AI bias. This includes finding and fixing biases in data and models, as well as raising awareness of how AI can impact human decision-making. By prioritizing fairness and neutrality in all data-driven processes, we can protect ourselves from the long-term harms of biased AI.

Types of Bias in AI

Data Bias:

  • Selection Bias: This occurs when the training data isn't representative of the real world. This can happen due to incomplete information, skewed sampling, or other factors that create an unbalanced dataset. Imagine a model trained only on data from male employees. It wouldn't be able to accurately predict performance for female employees.
  • Confirmation Bias: This bias arises when an AI system prioritizes pre-existing patterns or trends in the data too heavily. This can amplify existing biases and hinder the identification of new patterns.

Measurement Bias:

This bias occurs when the collected data differs significantly from the variables of interest. For example, a model trained to predict success in an online course based solely on data from students who completed the course might not accurately predict the performance of students who drop out.

Representational Bias:

  • Stereotyping Bias: This happens when an AI system reinforces negative stereotypes. This could be a facial recognition system with lower accuracy for people of color, or a language translation system that associates certain genders or stereotypes with specific languages.
  • Out-group Homogeneity Bias: This bias occurs when an AI system struggles to distinguish between individuals who are not part of the dominant group in the training data. This can lead to misclassifications or inaccuracies when dealing with minority groups.

Copyright

Current AI art models are trained on data that may include copyrighted material. Copyright laws may not be the best fit for all the challenges of AI. For example the Authors Guild argues that writers deserve credit and financial reward for their contributions to AI. They believe training AI models on copyrighted works is essentially copying that material. They propose a three-pronged approach:

  • AI companies should get permission to use copyrighted works in their programs.
  • Writers should be fairly compensated for both past and ongoing use of their work.
  • Writers should be compensated even if the AI outputs don't technically break copyright laws.

The Authors Guild believes collaboration with AI leaders is crucial to protect writers' livelihoods. They aim to unite writers and advocate for fair treatment within the realm of AI development.

The letter argues that generative AI wouldn't exist without the creative works of authors. Books, articles, and poems are used to train these AI systems, but authors aren't getting paid for their contributions. These works are like building blocks for language models. AI companies claim their machines just "read" the texts, but that's misleading. In reality, the texts are copied into the software and then used to create new content.

This puts the burden on artists to opt-out, rather than AI developers to ensure proper licensing. Some possible solutions may be:

  • Opt-in instead of opt-out: Require artists to explicitly allow their work to be used in training data.
  • Provenance tracking: Record details about the training data and generation process to improve transparency.
  • Audit trails: Create a system to track the origin and manipulation of data to avoid copyright infringement.

These improvements would benefit both AI developers and users by:

  • Reducing copyright infringement risks.
  • Enabling easier verification of AI-generated content.
  • Potentially allowing compensation for artists whose work influences AI outputs.

Here are some steps content creators and brands can take to protect their intellectual property (IP) from AI-generated content:

  • Monitoring - Content creators should use automated search tools to scan large datasets for their work (logos, artwork, text) that might have been incorporated into AI training data.
  • Proactive Defense - New tools can potentially obfuscate content from AI algorithms, making it harder to include in training data.
  • Trademark & Trade Dress Monitoring - Brands need to consider not just specific elements (logos, colors) but also the overall style of their work. This is because AI-generated content might not directly copy a brand's logo but still mimic its style, potentially infringing on trademark or trade dress.
  • Enforcement - The good news is that existing legal frameworks for trademark infringement still apply. Brand owners can use cease-and-desist letters, licensing demands, or lawsuits to address infringement, regardless of whether a human or AI generated the infringing content.

Hallucinations

AI hallucinations refer to situations where AI models confidently provide incorrect information that is not based on reality. This can be dangerous because users might not be able to distinguish between true and false information, especially for complex topics like court cases, health records, or secret company data. This means we shouldn't rely on these AI tools for tasks that require accurate information we can't immediately check ourselves. Because the internet is filled with untruthful information, these systems repeat the same untruths. A computer or AI application can’t tell the difference between truthful and untruthful data — all it sees is data. The core problem is that both human-entered and machine-derived data can be misleading for AI systems. 

Examples of AI hallucinations include:

  • Wrong answer about particular facts.
  • Creating fake legal cases.
  • Bad financial advice about compound interest.
Mistruths in Data

Understanding how data can be inaccurate is crucial. We'll examine five categories of mistruths in data

  • Mistruths of Commission (Lying): People provide false information instead of the truth (e.g., lying in an accident report). This can be intentional or unintentional.
  • Mistruths of Omission (Incomplete Truth): People tell the truth but leave out crucial details that change the story (e.g., mentioning everything in an accident report except texting while driving). This can be accidental due to forgetting or intentional.
  • Mistruths of Perspective: Different people see the same event from their viewpoints, leading to variations in the truth (e.g., driver, pedestrian, and witness have different accounts of an accident). AI struggles with this as it can only average the stories.
  • Mistruths of Bias: People are unable to see the truth due to personal beliefs or focus (e.g., driver focuses on the road and misses the deer). This can be difficult to identify and avoid.
  • Mistruths of Frame of Reference: People lack the experience to understand the information being presented (e.g., someone who has never been in a storm wouldn't understand the sailor's experience).

AI struggles with mistruths in data as it can not create its own experiences.

Privacy

AI's rise has fueled public excitement and shown us its potential to be beneficial. But it's also triggered important discussions on how to balance innovation with strong data privacy protections. In the US alone, 7 new state privacy laws passed in 2023, impacting AI development. While the future is uncertain, new privacy regulations and AI-specific laws are likely coming soon. As we navigate these changes, here are 3 key areas privacy professionals should focus on:

  • The use of public personal data in training models: We may need to redefine key privacy terms and update regulations entirely to keep pace with AI advancements.
  • Harmonizing privacy regulations with new AI regulation: AI regulations are on the rise globally, but how they'll work with existing privacy laws is a question mark. Governments need to consider how these new policies will interact in practice to ensure both ethical AI development and strong privacy protections.
  • Protecting children’s privacy: COPPA, the law protecting children's data online in the US, is getting updated. These changes are crucial as AI becomes more common in both products directly aimed at kids and those that might still interact with them. AI developers need to be mindful of how they use children's data and how their systems might interact with kids.
How to Use AI Tools Safely: Key Points

  • Always check the privacy policy: Before using any AI tool, understand what data they collect and how it's used.

  • Watch what you share: Avoid sharing personal or sensitive information with AI tools.

  • Change the settings: Most AI tools allow you to control how long your data is stored or to delete it entirely.

  • Specific actions:

  • Get proactive on privacy: Integrate privacy principles into your company's DNA from the beginning. This includes clear policies, employee training, and using tools like DPIAs and Google's Secure AI Framework to assess risks.
  • Build a privacy-aware culture: Everyone in the company should understand the importance of privacy. Encourage open communication about privacy concerns and empower employees to address them. Regular training and communication are key.
  • Leverage AI for compliance: Utilize AI tools like Checks to manage changing regulations, improve transparency, and respond efficiently to new requirements. AI can also provide a clear view of data usage, making compliance easier.
  • Consumer privacy concerns: Consumers worry that AI models trained on massive amounts of data, including web scraping and voice recordings, might infringe on copyright and misuse biometric data like voice prints. They fear this could limit their ability to make a living creatively and compromise their privacy.
  • Concentration of power: The vast amount of data needed to train AI models raises concerns about large companies gaining a competitive advantage through data aggregation and access to these powerful tools.

• Phishing scams: Generative AI could be used to write emails with perfect grammar and spelling, making them harder to distinguish from legitimate emails.

• Voice cloning scams: Scammers could use AI to clone the voices of loved ones for financial extortion.

• Romance scams and financial fraud: Chatbot products powered by AI could allow scammers to interact with more victims at once, increasing the effectiveness of these scams.

How to Spot Deepfakes: Key Points

The Detect Fakes experiment helps you learn about Deepfakes and how to identify them. Here are some key giveaways to watch for:

  • Focus on the face: Most Deepfakes manipulate faces, so pay close attention to this area.
  • Skin tone and aging: Look for inconsistencies between skin smoothness/wrinkles and the age of hair and eyes.
  • Eyes and eyebrows: Check for unnatural shadows around the eyes and eyebrows.
  • Glasses and lighting: Look for unrealistic glare or changes in glare as the person moves.
  • Facial hair: Examine the appearance and naturalness of facial hair, including additions or removals.
  • Facial moles: Pay attention to the realism of any moles on the face.
  • Blinking: Does the person blink at a normal rate or excessively?
  • Lip movements: In lip-synced Deepfakes, watch for unnatural lip movements.