In an ecosystem of over 100,000 active communities where context ranges from tactical video game advice to high-stakes political debate, evaluating user intent isn’t just difficult and it’s one of the most complex challenges in modern data science.
At the Nordic Data Science & Machine Learning Summit (NDSML), Alexander Gee, Staff Data Scientist at Reddit Inc., explained how Reddit transitioned from traditional transformer models to fine-tuned Large Language Models (LLMs) to keep users safe without silencing genuine conversation.
What does it take to moderate one of the world’s largest social platforms using state-of-the-art AI?
The Reason Why Out-of-the-Box Models Fail
When evaluating potentially harmful content, context is everything. Consider a simple controversial comment that involves weapons.
In isolation, that phrase triggers immediate safety red flags. But when posted in a sub-community dedicated to a challenging video game, it’s harmless gameplay advice.
Traditional automated systems and standard rule-based classifiers frequently trip over this exact nuance. When Reddit initially tested frontier LLMs out of the box, they encountered a different problem: over-policing.
Because commercial LLMs undergo alignment to make them ultra-cautious, they often stretch their interpretations of rules, flagging benign interactions as violations. Even aggressive prompt engineering (and “threatening” the model within the prompt) failed to achieve the high precision needed for automated enforcement.
Generic LLMs aren’t inherently aligned with your domain-specific tasks. High performance on public benchmarks rarely translates directly to production accuracy in niche operational environments.
Moving to Fine-Tuned Small Language Models (SLMs)
Rather than relying on massive, expensive commercial APIs, Alexander and his team shifted their approach toward training domain-tailored Small Language Models (SLMs).
By collaborating directly with Reddit’s operational experts and teams, known as anti-evil within Reddit’s territory, Trust & Safety built a high-quality data pipeline. Combining Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and parameter-efficient techniques like QLoRA, Reddit was able to:
- Teach smaller, faster models the specific nuances of Reddit’s community rules.
- Outperform state-of-the-art base models on internal benchmarks.
- Dramatically reduce inference costs while scaling detection capability to handle tens of millions of content removals annually.
Systems, Not Just Models: Building Real Guardrails
One of the biggest mistakes organizations make with LLMs is expecting 100% perfection from the model itself. In a live system, a single false positive can lead to an unjust account suspension and a broken user experience.
Reddit’s solution relies on a system-level guardrail architecture:
- SME Integration: High-ambiguity edge cases automatically route to human subject matter experts.
- Recourse & Appeal Loops: User appeals serve as real-time feedback loops, providing high-value labels to retrain and patch model blind spots.
- Layered Safety: Automated LLM decisions sit within a broader ecosystem of legacy automation, operational flags, and user reporting mechanisms.
By framing safety around overall system risk rather than isolated model metrics, Reddit successfully automated persistent edge cases that previously bottlenecked human moderation teams.
The Full Plan
This brief overview only scratches the surface of Reddit’s AI safety transformation. In the full session from NDSML, Alexander Gee shares the specific metrics, data labeling strategies, and architecture choices that made this deployment a success.
Learn how Reddit handles data quality auditing across billions of records, and discover actionable strategies for implementing QLoRA and DPO in your own data pipelines.
Become a Hyperight Member to Watch the Full Presentation
If you want to network with the minds behind similar systems and hear directly from the engineering leaders shaping the future of scalable, responsible AI, you can explore how to buy tickets for the upcoming summit.
