Generative AI introduces unique challenges beyond traditional AI systems. These models create new content—text, images, code, audio—which creates new types of risks and ethical considerations that organizations and practitioners must navigate.
What makes Generative AI different? GenAI is optimized to generate content, which expands the risk surface compared to classic predictive ML. A traditional AI might tell you whether an image contains a cat; Generative AI can create a brand new image of a cat. This creative capability brings novel challenges around truth, ownership, and control.
Hallucination is when GenAI confidently produces false or fabricated information.
| Type | Description | Example |
|---|
| Factual errors | Wrong information about the world | Inventing historical events |
| Citation fabrication | Fake or non-existent references | Made-up papers, URLs, people |
| Logical inconsistencies | Contradictory statements | Saying X, then later saying not X |
| Mathematical errors | Incorrect arithmetic or logic | Wrong calculations, flawed reasoning |
| Code errors | Non-functional or incorrect code | Wrong APIs, non-compiling code, insecure patterns |
| Cause | Explanation |
|---|
| Probabilistic nature | Models predict likely next tokens, not truth |
| Lack of grounding | Without retrieval/tools, models have no live fact source; with RAG/tools they can be grounded—if those components are correct |
| Weak calibration | Models can’t reliably distinguish known from unknown |
| Objective misalignment | Optimized for fluency, not accuracy |
| Strategy | How It Works |
|---|
| RAG (Retrieval-Augmented Generation) | Ground responses in retrieved facts |
| Human review | People verify outputs before use |
| Confidence signals | Can help but are not reliably calibrated; use with retrieval evidence and verification |
| Fact-checking layers | Separate verification step |
| Constrained generation | Limit responses to verified sources |
Practical tip: Treat GenAI as a creative assistant, not a truth-teller. Always verify important information.
GenAI models can perpetuate, amplify, or introduce biases present in their training data.
| Bias Type | Description | Impact |
|---|
| Representation bias | Under/over-representation in training data | Stereotypical outputs |
| Content bias | Skewed perspectives in source data | One-sided viewpoints |
| Objective/optimization bias | Loss function, RLHF, decoding push outputs toward certain styles | Certain outputs systematically favored |
| Deployment bias | Used in contexts different from training | Poor performance for some groups |
| Area | Bias Manifestation |
|---|
| Images | Stereotypical portrayals of professions, cultures |
| Text | Dialect or accent bias, cultural assumptions |
| Code | Western coding conventions, English-only comments |
| Recommendations | Perpetuates existing popularity disparities |
| Approach | What It Involves |
|---|
| Diverse training data | Include multiple perspectives, cultures, languages |
| Bias testing | Evaluate outputs across demographic groups |
| Fine-tuning | Adjust model for specific use cases |
| Red-teaming | Adversarial testing to find problematic outputs |
| Human oversight | Review and curation of training data and outputs |
GenAI models trained on copyrighted content raise complex legal questions that are still being resolved.
| Issue | Current Status |
|---|
| Training data copyright | Can you train on copyrighted content without permission? |
| Output copyright | Is AI-generated content protectable? |
| Style mimicry | Does imitating an artist’s style violate copyright? |
| Attribution | How do you credit training data sources? |
| Jurisdiction | Key Developments |
|---|
| United States | Copyright Office: AI-generated works not copyrightable without human authorship |
| European Union | AI Act: GPAI providers must publish summary of training content and comply with EU copyright rules (incl. opt-outs) |
| India | Active policy debate; proposals include licensing/compensation frameworks for AI training data (evolving) |
| Practice | Description |
|---|
| Opt-out mechanisms | Allow creators to exclude their work from training |
| ** licensing agreements** | Platform-specific licenses for AI training |
| Content credentials | Metadata indicating AI-generated vs. human-created |
| Model attribution | Documenting training data sources |
Guidance: If you’re using GenAI commercially, consult legal counsel. This area is rapidly evolving.
GenAI can be weaponized to cause harm at scale.
| Category | Risk Examples |
|---|
| Disinformation | Fake news, propaganda at scale |
| Social engineering | Convincing phishing, scam emails |
| Harassment | Targeted bullying, hate speech |
| Fraud | Deepfakes for financial gain |
| Cyberattacks | Automated vulnerability discovery, exploit generation |
| Non-consensual content | Deepfake explicit imagery |
| Challenge | Why It’s Difficult |
|---|
| Quality | High-quality AI content is hard to distinguish from human |
| Volume | Automated generation at overwhelming scale |
| Evolution | Techniques improve constantly |
| Plausible deniability | ”Real” vs. “fake” becomes harder to prove |
| Strategy | Approach |
|---|
| Watermarking | Embed signals indicating AI-generated content |
| Content authentication | Cryptographic provenance for media |
| Usage policies | Clear terms prohibiting harmful use |
| Red-teaming | Test for vulnerabilities before deployment |
| Monitoring | Track misuse patterns and respond |
Training large GenAI models requires significant computational resources.
Energy use varies widely with model size, hardware efficiency, tokens generated, and data center infrastructure. Estimates are highly uncertain.
| Aspect | Scale |
|---|
| Training (frontier models) | Third-party estimates: tens of GWh (high uncertainty; depends on compute, efficiency, re-runs) |
| Training (smaller models) | 0.1-1 GWh |
| Inference (per query) | Highly variable: fractions of a Wh to several Wh (depends on model, tokens, hardware, PUE) |
| Hardware | GPU manufacturing, data center infrastructure |
| Strategy | Impact |
|---|
| Efficient architectures | Smaller models, quantization, distillation |
| Renewable energy | Power data centers with green energy |
| Model reuse | Share models instead of training from scratch |
| Carbon tracking | Monitor and report energy consumption |
| Challenge | Description |
|---|
| Instruction following | Models may not obey constraints perfectly |
| Jailbreaking | Users can bypass safety measures |
| Prompt injection | Malicious instructions hidden in content |
| Objective misgeneralization | Model pursues narrow objective in unintended ways |
| Issue | Manifestation |
|---|
| Inconsistency | Different responses to the same prompt |
| Context window | Limited memory of conversation |
| Reasoning errors | Flawed logic in multi-step problems |
| Factual currency | Knowledge cutoff date limits |
- Hallucination: GenAI confidently produces false information; verify important outputs
- Bias: Models perpetuate training data biases; diverse data and testing help
- Copyright: Complex legal landscape; training on copyrighted content is unsettled law
- Misuse: GenAI enables new forms of abuse; watermarking and policies help
- Environment: Training is energy-intensive; efficient models and green energy help
- Technical: Control and reliability are ongoing challenges; human oversight remains essential
GenAI’s capabilities come with new responsibilities. Understanding these challenges is the first step toward mitigating them.