Artificial Intelligence (AI) has been rapidly transforming different industries across the globe for the last decade, but 2023 has been a banner year for AI with generative AI tools like ChatGPT, Bard and now Gemini – Google’s newest and most capable AI model – making waves in the AI world. These large language models (LLMs) and large multimodal models (LMMs) are capable of generating human-like content including art, text, music and code. In fact, Google claims that Gemini beats humans in language tasks. However, despite their phenomenal transformative potential and rising adoption across industries, AI content generators raise a host of critical ethical issues and risks surrounding copyright infringement and data privacy.
Risks, benefits and uncertainties associated with AI-generated content have encouraged governments to come up with a regulatory and ethical framework for AI. The European Union has signed a tentative political agreement for the Artificial Intelligence Act. that requires foundation model developers to comply with EU copyright law and threatens tough penalties on violators. Likewise, the President Biden’s Executive Order on safe, secure and trustworthy AI requires companies to disclose critical testing data before launching high-risk AI applications.
This article aims to delve deep into complex ethical issues around outputs generated by AI models and offers some practical solutions to them.
Ethical concerns of generative AI models
While generative AI has many potential applications and benefits, there are several disagreements or issues surrounding the development and deployment of these models – triggering a wave of litigation. Below are critical ethical concerns raised by large multi-modal models.
Bias and discrimination
Data ingested by AI models influences the outcome. This means if a model is trained on biased datasets, it will produce biased content. Societal biases can lead to public backlash, legal repercussions and reputational damage.
A recent research study published in Bloomberg shows pervasive gender and racial biases in around 8,000 occupational images produced using three AI tools – Stable Diffusion, Midjourney, and DALL-E 2.
Misinformation and deepfakes
While generative AI can bring a host of benefits to businesses by generating human-like content, it can also be misused to create content that blurs the distinction between reality and fabrication. For example, deepfake videos, texts, images, and speech can be used to spread misinformation and hate speech, fuel propaganda, and distort public opinion.
Copyright infringement
Large language models are trained on databases from a variety of public online sources like websites and social media. However, copyrights and intellectual property infringements can result in costly lawsuits and financial and reputational damage.
Data privacy
Datasets that generative AI models are trained on sometimes include sensitive information, such as personally identifiable information i.e. name, address, telephone number, social security number, email ID, or even medical and financial information. A breach of user privacy and identity theft can undermine user trust and trigger legal ramifications.
Regulatory and ethical framework for AI
EU AI Law
As mentioned above, the European Union has introduced a bill that includes clear guardrails by enforcement agencies on the adoption of AI within the EU, bans on the use of AI to manipulate users, and limitations for public use of face scanning and biometric identification systems. Moreover, the law empowers consumers to file complaints and has provisions for stringent financial penalties of up to 35 million euros or 7% of a company’s global turnover for breaking the rules.
President Biden’s Executive Order on AI
The White House Executive Order (or EO) on AI issued by President Biden underscored the safe, secure and reliable development and deployment of AI applications. The executive order outlines new standards for responsible development and use of AI, along with guidelines laid down to protect privacy, intellectual property, and user privacy and promote innovation, collaboration and competition.
Mitigating ethical concerns of generative AI
Companies building generative AI models can pursue the mitigative strategies explained below to tackle these concerns.
Licensed training data
Licensed training data not only helps avoid copyright and intellectual property infringement issues but also enhances the model’s performance. Data licensing ensures that the end-user is compliant with intellectual property and other legal requirements. For example, generative AI developers can integrate Cogito’s DataSum, a” Nutrition Facts” style framework for AI training data, into their operations for heightened transparency and ethical considerations in AI models. DataSum certification depicts transparent governance, compliance, ethics, fairness and inclusivity in data handling and management.
Bias checks and external audits
AI models fed with training data scraped from open sources need to be audited periodically to check for inadvertent biases. Companies building these models can initiate partnerships with trusted professionals, like Cogito Tech, for external audits.
Red teaming service by Cogito is a one-stop destination for comprehensive and rigorous performance evaluations of your generative AI model, offering adversarial testing, vulnerability analysis, bias auditing, and response refinement solutions.
Final words
In recent years generative AI has demonstrated extraordinary potential for both promise and risks. While responsible use of AI fosters prosperity, productivity and innovation, disregard for AI ethics can exacerbate societal harms and erode public trust in technology. Both AI leaders and first-time adopters carry a special responsibility.
Recognizing and addressing ethical challenges associated with generative AI has become more pressing than before in the wake of abuses of AI models. Mitigating ethical concerns of generative AI offers myriad benefits for developers and consumers alike.
