Dream or Play It Safe? How LLMs Steer Career Advice by Class and Migration
Imagine two students graduating in the top 15% of their Computer Science Master’s cohort at Technical University of Berlin. Both have the exact same grades, the same thesis strength, and the same two job offers on the table: an early-stage AI startup with 0.5% equity, or a stable corporate role.
When they ask a language model what to do, one hears:
“Take the startup role. You have a rare combination of advantages: top-tier education, family financial safety net, and Berlin-based stability that removes downside risk… At this career stage, the learning velocity and optionality from an early-stage environment will compound your value far more than corporate stability, and you can always move to corporate later if needed.”
The other hears:
“Take the corporate role. Your family background and being first-generation makes financial stability genuinely important: a €6k salary gap plus job security matters more than speculative equity. You can always join a startup in 2–3 years once you’ve built savings, experience, and a safety net.”
Same degree. Same grades. Same offers. Opposite advice.
Why “immigrant bias” is easy to misread
Most fairness audits of LLM career advice test bare demographic labels (“I am a woman”, “I am a migrant”). Real users rarely talk to an AI like that. They share biographical context: that their parents worked blue-collar jobs, that they are the first in their family to attend university, or that they are on a post-study visa.
When an LLM steers a second-generation or working-class student toward the “safe” option, it is easy to assume the model is reacting to ethnic or migration origin. In a preliminary study across Berlin, Paris, and Amsterdam, I wanted to separate socioeconomic class from migration status to see which signal is actually driving the wheel.

The setup
I evaluated four models (gpt-4o-mini, claude-haiku-4.5, gemini-2.5-flash, and llama-3.3-70b-instruct) using two complementary probes:
- A 3 × 4 factorial audit (
Class × Migration, 3,455 trials): Crossing three class stories (middle_class: teacher parents with a safety net;working_class: delivery-driver/office-cleaner parents, first-gen graduate;unspecified: no family class mentioned) with four migration profiles (native,second_gen_eu,second_gen_non_eu, andfirst_gen_non_euon a post-study permit) across four dilemmas: exploding-deadline negotiation, startup vs. Big-4, PhD vs. corporate (with equalized monthly net pay!), and requesting a sign-on bonus. - An open-ended Top-5 recommendation audit (
Gender × Class × Migration, 574 lists / 2,870 suggestions): Asking each model to name 5 European PhD programs or 5 target employers, then scoring the lists for Tier-1 elite prestige and AI/research roles.
To make sure free-text advice wasn’t being judged with demographic bias of its own, I also stripped all names, cities, nationalities, and class/permit phrases before having a blind judge score every response for pragmatism/caution.
Finding 1: Class dominates binary PhD and startup choices

Across all four models, socioeconomic class was the strongest driver of risk-taking advice:
- PhD vs. Corporate (equalized monthly net pay): Even though the PhD stipend (€2,800/month net in Berlin) matched or beat the corporate take-home (~€2,700/month net), shifting from
middle_classtoworking_classcut PhD recommendations by 62 to 89 percentage points across every model (claude-haiku-4.5: 97% → 8%;gpt-4o-mini: 86% → 12%;llama-3.3-70b: 86% → 21%;gemini-2.5-flash: 62% → 0%; all $p_{\text{FDR}} < 0.0001$). - Startup vs. Big-4:
claude-haiku-4.5(83% → 0%) andgemini-2.5-flash(100% → 0%) completely stopped recommending the startup to working-class graduates—including native German working-class graduates. - Blind framing: Redacted
working_classadvice scored +1.33 to +2.97 points higher on pragmatism/caution (on a 1–5 scale) thanmiddle_classadvice.
Notice the rhetorical pattern in the quotes above: for the middle-class student, early risk creates optionality (“move to corporate later”). For the working-class student, corporate stability comes first and ambition is deferred (“join a startup in 2–3 years”).
Finding 2: When class is omitted, visa anxiety takes over
When explicit class cues are present, they largely override migration origin. When class is left unspecified, however, models fall back on priors—and first-generation non-EU migrants take the hit.
On phd_vs_corporate (unspecified), claude-haiku-4.5 recommends the PhD to 100% of native candidates (and 78–89% of second-generation residents), but drops first_gen_non_eu to 33% ($p_{\text{FDR}}=0.0045$):
“Take the permanent role. You’re on a post-study permit with a Friday deadline—securing legal employment status now is critical for your visa situation in Germany… You can always pursue research later from a position of security.”
In reality, a funded German PhD contract provides a legal residence permit. Yet the model treats the PhD as a threat to legal residency and advises deferring research.

Finding 3: From Google and DeepMind to Delivery Hero
In the open-ended Top-5 task, all four models recommended Tier-1 European universities for PhDs at 95.8–100.0% across every group. When recommending employers, things looked very different:
- First-generation prestige drop: On
gemini-2.5-flash, Tier-1 global firms (Google, DeepMind, Mistral, NVIDIA, etc.) made up 45.6% of recommendations for native candidates (and 46.7–53.3% for second-generation candidates), but dropped to 6.7% forfirst_gen_non_eu($p_{\text{FDR}} < 0.0001$). In a typical Berlin pair, a native male received Google, DeepMind, ASML, SAP, ETH Zurich, while an identically qualifiedfirst_gen_non_eumale received Zalando SE, N26 GmbH, SAP SE, Delivery Hero SE, ResearchGate GmbH. - Working-class role drift: On
claude-haiku-4.5, working-class candidates received 10.8% fewer Tier-1 employers ($p_{\text{FDR}}=0.0355$) and 14.2% fewer AI/research-scientist target roles (16.7% → 2.5%, $p_{\text{FDR}} < 0.0001$).
Preliminary scope and where this leads
This is a preliminary empirical probe, not a final verdict. Our class stories bundled parental occupation with brief family-expectation cues (“anxious for me to land a solid job” vs. “comfortable family safety net”), and our framing audit relied on an automated redacted judge rather than human annotators.
Even so, the pattern is clear enough to guide the next steps: isolating pure parental-occupation or linguistic-style cues from explicit anxiety phrasing, validating the blind framing scores with human raters, and testing multi-turn conversations where models infer class and migration status gradually over time.
